Cortical Correlates of Low-Level Perception: From Neural ... · physiological data describing the...
Transcript of Cortical Correlates of Low-Level Perception: From Neural ... · physiological data describing the...
Neuron
Perspective
Cortical Correlates of Low-Level Perception:From Neural Circuits to Percepts
Yves Fregnac1,* and Brice Bathellier11Unite Neuroscience Information et Complexite (UNIC), CentreNational de la Recherche Scientifique, FRE 3693, 91198Gif-sur-Yvette, France*Correspondence: [email protected]://dx.doi.org/10.1016/j.neuron.2015.09.041
Low-level perception results from neural-based computations, which build a multimodal skeleton of uncon-scious or self-generated inferences on our environment. This review identifies bottleneck issues concerningthe role of early primary sensory cortical areas, mostly in rodent and higher mammals (cats and non-humanprimates), where perception substrates can be searched at multiple scales of neural integration. We discussthe limitation of purely bottom-up approaches for providing realistic models of early sensory processing andtheneed for identificationof fast adaptiveprocesses, operatingwithin the timeof apercept. Futureprogresseswill depend on the careful use of comparative neuroscience (guiding the choices of experimental models andspeciesadapted to thequestionsunder study), on thedefinitionof agreed-uponbenchmarks for sensory stim-ulation, on the simultaneous acquisition of neural data at multiple spatio-temporal scales, and on the in vivoidentification of key generic integration and plasticity algorithms validated experimentally and in simulations.
IntroductionLow-level perception can be defined as the neural-based com-
putations building unconscious or self-generated inferences
during the processing of sensory events. These internal infer-
ences are in turn ‘‘psychologically projected into external space
and accepted as our most immediate reality’’ (Gregory, 1997).
Low-level perception does not necessarily require attentional
processes, and, for this reason, it is often confounded with
non-attentive perception. This operation constitutes a necessary
reformatting of sensory input in a compositional neuronal lan-
guage, which will allow more abstract processing by higher
cognitive-like cortical areas (see Abboud et al., 2015). This
grammar of perceptual primitives, which emerge (‘‘pop out’’)
effortlessly, reflects the expectations of our brain, built through
prior sensorimotor experience, about what is to be perceived
(Gregory, 1997). This largely autonomous, dynamic, and adap-
tive process can be observed in the anesthetized brain, although
it is continuously updated through closed-loop interactions in the
awake and attentive states. It feeds and guides our cortically
mediated interactions with the world through fast, but context-
adaptive, behavioral choices (Figure 1A).
Perception is based on raw information detected by sensory
receptors of the organisms, usually classified into distinct
‘‘senses’’ or modalities. Following Aristotle, five major senses
are classically distinguished in humans, depending on the type
of physical signal used for transduction (photon for vision, acous-
tic vibration for hearing, volatile anddissolved chemicals for taste
and olfaction, mechanical forces for the haptic sense). Sherring-
ton later proposed a more functional classification based on the
spatial proximity of the sensory sources relative to our own
body: vision, olfaction, and hearing are the ‘‘teleceptors’’ inform-
ing about the distant environment, while touch and taste are the
‘‘exteroceptors’’ informing about what is within our immediate
reach; in addition, ‘‘interoceptors’’ inform about the bodily func-
tions and ‘‘proprioceptors’’ inform about bodily position.
110 Neuron 88, October 7, 2015 ª2015 Elsevier Inc.
A key attribute of perception is that it usually solves ambigu-
ities that result from incomplete data about the world. Since
ancient debates opposing rationalists and empiricists, it has
been clear that ‘‘what we see (perceive) is not what we get
(receive)’’: perception departs from the physical reality sampled
by our sensors, leading to illusions that reflect ‘‘erroneous’’ infer-
ences about the sensory input. Figures 1B and 1C illustrate two
classical examples of illusions: the Mach bands illusion (Ratliff,
1965), in which discontinuities in the spatial derivative of lumi-
nance induce the perception of illusory contrast edges, and the
‘‘line motion effect,’’ in which the sequential presentation of
two static shapes (a square followed by a rectangle) produces,
for a specific range of stimulus intervals, the perception of
continuousmotion (‘‘phi’’ effect inWertheimer, 1912). Aswe later
develop, canonical circuit motifs forming built-in physiological
filters can account for such biases of visual perception.
As it has long been noticed, a complementary way of solving
ambiguity in object identification is to combine or transfer
perceptual knowledge across sensory channels. In his famous
letter to John Locke in 1688, William Molyneux asked whether
a born blind man who has learnt to distinguish a globe and a
cube by touch would be able to distinguish and name these ob-
jects simply by sight, once he had been enabled to see. Recent
observations in newly sighted humans indicate that the answer is
no, but that cross-modal associations can be rapidly formed
(Held et al., 2011). Hence the use of contextual information in
perception through ‘‘built-in’’ relationships may generalize from
unimodal to multimodal perception. Multimodal illusions support
this idea. For example, in the double flash illusion, a tone triggers
the percept of a luminance change (Watkins et al., 2007), while in
theMcGurk-McDonald effect, the percept of a vocalized syllable
is biased by the visual observation of mismatched mouth mo-
tions (McGurk and MacDonald, 1976). Multisensory integration
is also at the center of our everyday experience of ‘‘body-owner-
ship’’ (see also Blanke et al., 2015). The representation of body
Figure 1. Sensation, Perception, and Illusion(A) A conceptual framework (inspired by Richard Gregory).(B) Mach bands: from top to bottom, displayed luminance profile, measured luminance gradient in positions 1–4, perceived luminance profile. The Mach illusion(blue arrows) corresponds to the perceived lighter (position 2) and darker (position 3) bands.(C) Line motion illusion: top, spatio-temporal sequence of displayed stimuli (square followed by a bar); bottom, perceived downward motion (blue arrow). (A)–(C)adapted from Shulz and Fregnac, 2010.(D)Pictureof the surfacevasculatureoverlying the imaged regionofV1,V2, andV4with the regionof interest (ROI), alongwithdiagramsof the visual stimuli used tomapcortical responses. Scale bar = 1mm. The red arrows indicate the drift directions of the stimuli. Red outlines delineate the areas of V1, V2, and V4 that were analyzed.(E) Differential analysis of the orientation maps recorded simultaneously across the three different visual areas: response profiles produced by gratings (LG) andillusory contours (IC) were indistinguishable in V4, while those in V1 and V2 showed a clear orientation preference shift for IC stimuli. The observed shift reflects the45� angle between the solid inducer orientation and the illusory contour.(F) The relative scale of spatial integration of visual cortical receptive fields is overlaid with the pattern of illusory contours (IC) defined by spatially aligned abuttinghorizontal lines. The hierarchical functional hypothesis of Wang and colleagues is that spatially aligned neurons with smaller RFs in V1 and V2 project to IC-selective neurons with larger RFs in V4, therefore generating a more robust response to IC in V4. For illustrative purposes, the RF diameter of V1 and V2neurons was taken as 1� and that of V4 neuron as 6�. The mechanisms regulating the coordinated activity between feedforward and feedback flow are stillunknown. (D)–(F) are freely adapted from Pan et al., 2012, with permission of Wei Wang.
Neuron
Perspective
parts is intrinsically multimodal, as revealed by illusions such as
the rubber hand illusion (Botvinick and Cohen, 1998), in which
the pairing of a tactile stimulus on the real hand of a subject
with visualization of a touch on a fake hand creates an ownership
illusion that the fake hand belongs to the body. This is also partic-
ularly evident for the global consciousness of one-self as an
entire body, which can be profoundly affected in situations in
which, as in the above example, incongruent multimodal evi-
dence is given to the subject (Blanke, 2012).
To restrict the scope of this article, we will focus on the role
played by primary sensory cortical areas in low-level perception,
by taking examples in a few animal species where the substrates
of perception can be looked for at multiple scales of neural inte-
gration, mostly in the rodent and higher mammals (cats and non-
human primates). David Marr (Marr, 1982) distinguished three
hierarchical levels of analysis in the study of neural systems: (1)
the computation realized by the brain, (2) its algorithmic instanti-
ation, and (3) the biophysical substrate of implementation. Start-
ing from the psychological or psychophysical evidence, which
allow us to characterize the computation corresponding to
low-level perception, we will review current knowledge on the
elementary algorithmic principles found in cortical circuits un-
derlying perception (see Box 1) and discuss the main bottleneck
issues for extracting the rules necessary to build any realistic
model of the early sensory brain (Box 2).
This review will draw mainly from unimodal perception in
vision, which offers a unique and abundantly documentedmodel
of sensory perception necessary for higher cognitive activities
(e.g., sign communication, face recognition, and reading and
writing in humans) and structurally prevalent in the neocortex
of higher mammals (more than half of primate cortex is allocated
to the analysis of visual information distributed across several
tens of individualized retinotopically organized areas; Markov
et al., 2013). Moreover, important advances have been made
on the more holistic aspects of high-level vision (see Abboud
et al., 2015), which we could start to bridge to the abundant
physiological data describing the elementary mechanisms at
play in the early (subcortical and cortical) visual system.
Phenomenology of Low-Level Perception: A HolisticPerspectiveA natural starting point for characterizing perception is to ask:
how are sensory ‘‘bits’’ bound in a ‘‘whole’’ that is meaningful
to the observer?When facing a natural visual scene, human sub-
jects have an immediate conscious perception of the elementary
features that compose it (segmentation), as well as of the higher-
Neuron 88, October 7, 2015 ª2015 Elsevier Inc. 111
Box 1. Current Status of the Field
d ‘‘Pop-out’’ non-attentive perception is constrained by ho-
listic laws that reflect expectations about the composi-
tional structure of what is to be ‘‘perceived.’’
d No definitive neural-based bottom-up model accounts for
the emergence of these psychic laws.
d Models of receptive fields fail to consider the synaptic
nature of their underlying processes and consequently
underestimate the impact of local recurrent networks.
d Models of receptive fields fail to account for the functional
diversity of contextual dependencies and for larger-scale
interactions across features.
d Presentmodels of perception rely on a dominantly feedfor-
ward hierarchy of cortical areas, but multiple examples
attest for the backpropagation of higher-order feature
selectivity onto primary sensory cortical areas.
d Algorithms of fast and reversible plasticity necessary
for transient percept emergence remain to be validated
in vivo.
d The jump from rodent to humans regarding the neural or-
ganization of low-level perception seems presently too
high.
Box 2. Future Directions
d Amathematical framework that goes beyond the receptive
field concept is necessary to define the complex nonlinear
transformation applied by brain circuits tomap physical re-
ality into the perceptual space
d Multiscale models that realistically account both for struc-
tural and functional observations will help to access the
algorithmic principles of cortical circuits.
d Large-scale recordings during uni- and multimodal
perception will make it possible to test if specific attracting
states of cortical neuron populations represent elementary
tokens of percepts.
d New techniques to precisely control spatio-temporal acti-
vation sequences in cortical assemblies will help to eluci-
date the exact nature of the neural code and the temporal
precision below which the percept is lost.
d Comparative physiology approaches should be encour-
aged to discriminate the common principles shared by cir-
cuit motifs from a species-specific use of a sensorymodal-
ity or multimodal strategy.
Neuron
Perspective
order global objects that emerge from their associations (bind-
ing), although not necessarily in this order. Contours, colors,
textured surfaces, shapes, and three-dimensional objects pop
out unambiguously, in a fraction of a second to seconds, accord-
ing to the background or context.
Various conceptual schools have emerged from the holistic
approach. The first attempt came from Gestalt theory, intro-
duced by von Ehrenfels and generally attributed to Kolher. In
its simplest form, Gestalt theory posits that the ‘‘whole’’ and
the ‘‘parts’’ of a percept coexist in a reciprocal dependency
that obeys specific relational rules, such as proximity, similarity,
uniform density, common fate, direction, good continuation,
closure, and symmetry, to name but a few (for review seeWage-
mans et al., 2012). Gestalt psychology remains relevant to our
current understanding of the emergence of relational structure
and higher-order binding. By definition, Gestalt effects are illu-
sion makers, but these illusions are the expression of our brain
grouping information and building, according to Aristotle’s
view, global constructs ‘‘other than the sum of their parts.’’ A
fascinating illustration of the compositional power of perception
is given by the simultaneous extraction, within a single glance, of
both local (fruit) and global (face) features, from Arcimboldo’s
portrait of Emperor Vertumnus (the Roman God of the Seasons).
The original claim of the ‘‘Gestaltists’’—the existence of a
‘‘grammar of seeing’’ as phrased by Kanizsa—has been criti-
cized and extended by other schools of thought. The field of ‘‘en-
action’’ (Palacios and Bozinovic, 2003), for example, insists on
the learnt nature of the elements of this grammar, reflecting
active perception and interactions with the environment, as
opposed to fixed perceptual primitives reflecting the structure
of objects. Other approaches focused on mathematical formula-
tions of this idea: e.g., artificial vision models (Morel et al., 2010)
or symbolic modeling (for review see Wagemans et al., 2012).
112 Neuron 88, October 7, 2015 ª2015 Elsevier Inc.
Nevertheless, there is a consensus that key perceptual illusions
reveal the existence of perceptual laws, characterizing the pro-
cess of recognition of objects, and that these illusions can be
used to reveal the neural correlates of perception and its compu-
tational underpinnings (Eagleman, 2001).
Several visual illusions in which the specificity of the output is
not predicted by that of the ‘‘parts’’ and their topological union
provide other important clues. For example, binocular rivalry, in
which a different image, or different non-corresponding image
patches (Diaz-Caneja, 1928) are simultaneously presented to
each eye, results in random switching between two possible per-
cepts. This paradigm illustrates the holistic nature of the percep-
tual process by providing evidence for a single percept among
several possible under multiple conflicting inputs. Moreover, it
enables the neurobiological study of how this percept emerges.
Studies performed at the single-cell level in the behaving non-hu-
man primate indicate that the percentage of visual cortical neu-
rons exhibiting activity correlated with the perceptual report in-
creases along the visual cortical area hierarchy, starting with a
few ‘‘reporters’’ in V1 and culminating in the inferotemporal cortex
and the superior temporal sulcus (Leopold and Logothetis, 1996).
Most remarkably, unilateral hemisphere activation (by a caloric
stimulus applied to one inner ear or by transcranial magnetic
stimulation) forces the alternation between the reported rivaling
percepts, which suggests that the conscious percept is selected
through competition between the two hemispheres (the ‘‘inter-
hemispheric switch hypothesis’’ (Miller et al., 2000). These results
suggest that the ambivalence is present in the information distrib-
uted across the sensory cortical hierarchy and that only one of the
two encoded percepts is raised to awareness through a switch
operating at the scale of interacting networks (mesoscopic scale).
The emergence of a global percept together with its corre-
sponding features implies that the brain brings certain features
(potentially even ones absent from physical reality) to conscious-
ness and discards others. For example, in the tactile funneling
Neuron
Perspective
illusion, simultaneous stimulations of two adjacent fingers evoke a
single focal sensation at the center of the stimulus pattern (central
to the two digit tips or spanning the two stimulation sites), even
when no physical stimulus occurs at that site (Chen et al.,
2003). In the ‘‘line-motion’’ illusion (Figure 1C), a stationary square
briefly precedes a long stationary bar presentation and produces
a perception of continuousmotion, absent in the stimulus (Jancke
et al., 2004). In the case of subjective or illusory contours (ICs;
Figure 1D), the emergence of ‘‘non-existent’’ features (contours)
is induced by precise geometrical alignment of real stimulus fea-
tures such as abutting lines (Pan et al., 2012; Seghier and Vuil-
leumier, 2006). These illusions show that the brain does more
than mirroring retinal inputs; it also computes knowledge or infer-
ences encoded in the processing circuits, such as the continuity
of objects boundaries or displacements. Extensive studies
show the cortical origin of these effects, at many levels. Simple
fusion illusions such as the ‘‘funneling’’ and ‘‘line-motion’’ effects
are already reflected in the activity of V1, as shown by voltage-
sensitive dye imaging in monkey and cat (Chen et al., 2003;
Jancke et al., 2004). These may be accounted for by lateral diffu-
sion processes intrinsic to primary sensory areas (Fregnac, 2012).
By contrast, the emergenceof illusory contours (as in Figure 1D)
involves reverberation across multiple networks including sec-
ondary visual areas. fMRI studies in humans have demonstrated
robust activations to illusory contours in V1 and V2 (for review,
see Seghier and Vuilleumier, 2006), but also distinct activations
in areas such as V4 (Mendola et al., 1999; Montaser-Kouhsari
et al., 2007). Electrophysiology in animal models shows a propor-
tional increase in the number of visual cells responding to illusory
contours between V1 and V2 (Lee and Nguyen, 2001; von der
Heydt et al., 1984), but these responses occur a few tens of milli-
seconds later than in downstream areas such as V4 (De Weerd
et al., 1996; Lee and Nguyen, 2001), suggesting that they may
reflect ‘‘top-down’’ cortico-cortical feedback. A recent imaging
and electrophysiology study (Pan et al., 2012) in rhesus ma-
caques, confirming partially a pioneer study in cat V1 and V2
(Sheth et al., 1996), shows that the orientation preference maps
for drifting gratings (DG) and virtual contours (VC) are shifted by
the angle between the solid inducer and the virtual contour in V1
and V2, whereas they superimpose perfectly for V4 (Figures 1D–
1F). Electrophysiological population data confirmed a ‘‘salt and
pepper’’ organization of VC-responsive cells in V1 and V2 already
seen by others. Thus, the orientation domains mapped in early vi-
sual areas V1 and V2mainly encode the local physical features of
the inducer stimulus, whereas a complete overlap of orientation
domains to process real and illusory contours emerges only in
V4. The roleofhigher-orderareas in thisphenomenon issupported
by lesion studies, showing that the perception of ICs, but not of
luminance-defined real contours, is severely impaired after V4
lesion (DeWeerd et al., 1996). In summary, the global organization
of virtual contour decoding seems to agree (superficially) with a
global and distributed multilayered feedforward hierarchy culmi-
nating in V4, but its functional emergence obviously requires the
concomitant activation of cortico-cortical feedback processes in
V1, through mechanisms which are far from being elucidated
(Figure 1F).
Similar back-propagation of the functional influence of higher-
order selectivity onto primary sensory cortical areas is seen also
in multiple context-modulation effects (Sharma et al., 2003; Su-
per et al., 2003; Zipser et al., 1996) and during multimodal
discrimination task learning (Lemus et al., 2010).
The idea of a ‘‘perceptual grammar’’ shaping sensory informa-
tion according to defined compositionality rules extends from
uni- tomultisensory processing. The ventriloquist effect, in which
vision biases sound source localization, provides a good ex-
ample. Just as the facial motion of a puppet expertly animated
by a ventriloquist gives us the illusion that the puppet speaks,
the coincident appearance of an object (even a flashed circle)
in our field of view together with a sound often leads to the incor-
rect impression that the sound originated from the location of the
visual stimulus (Jack and Thurlow, 1973). This effect suggests
the expectation (innate or learned) of a spatial correlation be-
tween inputs, reflecting physical reality. The magnitude of the
cross-modal effect depends on factors such as spatial (Jack
and Thurlow, 1973) and temporal proximity (Bonath et al.,
2014) of the stimuli. Functional electroencephalographic and im-
aging studies indicate that cortical areas traditionally seen as
unisensory (such as the core/belt region of the auditory cortex)
are, together with associative areas, involved in the ventriloquist
illusion (Bonath et al., 2014). Thus, multisensory integration ap-
pears to engage multiple levels of cortical interactions, possibly
across the entire hierarchy of sensory areas.
Bottleneck Issues
Even if holistic models are conceptually rich and powerful, they
remain difficult to link to precise explanatory neural and synaptic
mechanisms, especially via unsupervised means. Except for
Grossberg’s attempt to build a consistent neutrally inspired
computational framework of early visual processing (Grossberg
et al., 2008), entirely bottom-up approaches have not yet pro-
duced convincing implementations of Gestalt principles. This
failure was, in a way, predicted by Wertheimer in his definition
of holism: ‘‘There are wholes, the behavior of which is not deter-
mined by that of their individual elements, but where the part-
processes are themselves determined by the intrinsic nature of
the whole.’’ Although often opposed, Gestalt theory (calling for
a holistic evaluation) and neural-based theories of object recog-
nition are not mutually exclusive: when driven dominantly by
external sensations and bottom-up activation, the cortical
neuronal machinery generates an automatic interpretation of
our environment through built-in mechanisms yet to be identi-
fied, whose effects contribute to the emergence of perceptual
laws applicable to automatized sensing. When driven mainly
by top-down expectations, the computation takes advantage
of a zoo of hallucinatory states internally stored in the same net-
works, to extrapolate what is given to be seen (‘‘controlled hallu-
cination,’’ as coined by Jan Koendrink [Koenderink, 2012]).
Turning these ideas into working models is still beyond reach,
but hopefully possible. Below, we will try to identify the gaps
that need to be filled to attain this goal.
Re-evaluation of the Receptive Field ConceptEfforts to characterize, in cortical sensory areas, the algorithms
of low-level perception (as in Marr’s tripartite definition) focused
initially on the attributes and perceptual invariants encoded by
single neurons. The dogma was that a neuron encodes the input
configuration that maximizes its firing frequency (the ‘‘neuronal
Neuron 88, October 7, 2015 ª2015 Elsevier Inc. 113
Neuron
Perspective
doctrine’’ of Barlow [Barlow, 1972]). This paradigm naturally
extended to the receptive field (RF): a neuron’s RF can be
defined as the region of sensory space within which the pres-
ence of a stimulus significantly modifies its activity (Hartline,
1938). In the visual system, the RF is the region of the retina
where a local change of luminance (relative to the background)
modulates (enhances or suppresses) the firing rate or the prob-
ability of discharge of the recorded neuron. In the auditory
system, RFs correspond to regions in physical space, in the tem-
poral and frequency domain, or both. In the somatosensory sys-
tem, receptive fields are (usually contiguous) regions of the body
surface or internal organs. A special case of tactile receptive
fields are those corresponding to the whiskers in rodents. This
operational concept used across most sensory modalities (but
see below) has provided a powerful description of neuronal re-
sponses in the visual system with an emphasis on primary visual
cortex (Alonso and Swadlow, 2005) and area MT (Rust et al.,
2006), in somatosensory cortex (Estebanez et al., 2012), and in
auditory cortex (Theunissen and Elie, 2014).
Although receptive fields are primarily a convenient way to
categorize the response of neurons, they imply the existence
of hierarchies of representations in which each neuron codes
for a particular feature of the physical space in a static and
invariant way and in which a scene can be reconstructed in a
predominantly ‘‘bottom-up’’ fashion from assemblies of active
neurons and the knowledge of their associated features. In
particular, in the early visual system of mammalian carnivores
and primates, one observes an undeniable progression in RF
complexity: isotropic ‘‘Mexican hat’’ receptive fields with con-
centric opponent ON and OFF subfields, found in the retina or
the thalamus, detect local contrast change and filter out back-
ground mean luminance. In V1, simple RFs with spatially segre-
gated ON and OFF subfields are selective to position and
contrast edge orientation, while complex RFs generalize orienta-
tion selectivity across position within the receptive field. In V2
and V4, higher-order hypercomplex cells seem to code for cor-
ners and masking singularities. Even if this simplifying scheme
has been challenged, it has inspired the hierarchical organization
of the most advanced models of shape recognition (DiCarlo
et al., 2012) and provided adequate component filters to extract
a perceptual sketch of physical shapes. Hence, piecewise RF
decomposition is a powerful tool to describe the low-level skel-
eton of sensory percepts across subcortical and cortical areas.
This formalism is not without its problems, however. For
example, the full description of a RF requires, in principle, that
one tests the entire space of possible stimuli on a neuron; this
is clearly impossible, especially with poorly parameterized stim-
ulus spaces (e.g., chemicals in olfaction, natural scenes in
vision). In practice, people make simplifying assumptions: for
example, that a RF can be reduced to a linear filter combined
with a static nonlinearity (LNP models; Figures 2A and 2B)
of varying complexity (e.g., Priebe and Ferster, 2012). This
assumption enables the use of reverse correlation techniques
with broadband stimuli (e.g., white noise) to derive the filter—a
process whose logic is obviously circular. Observations in pri-
mary visual cortex (Baudot et al., 2013; David et al., 2004; Four-
nier et al., 2011; Smyth et al., 2003), auditory cortex (Bathellier
et al., 2012; Machens et al., 2004), and even in the retina (Hosoya
114 Neuron 88, October 7, 2015 ª2015 Elsevier Inc.
et al., 2005) indicate that linear filters with a static nonlinearity
poorly predict neuronal responses, especially for natural stimuli.
Linear receptivefielddescriptionscouldbeextendedwith ‘‘gain
control’’ mechanisms in retina (Solomon et al., 2006), thalamus
(Mante et al., 2008), and primary visual cortex (Rust et al., 2005),
which can have low-contrast enhancement functions as the ‘‘divi-
sive normalization’’ principle (Carandini and Heeger, 2012) in
which the output response of a neuron is divided by a factor pro-
portional to the average (or the sum) firing rate of nearby neurons.
A further, mathematically more general refinement is to
describe the RF as a bank of parallel linear and nonlinear sub-
units, using two main approaches that are, to a certain extent,
equivalent mathematically: spike-triggered covariance analysis
(Rust et al., 2005) (Figure 2C) or Volterra decomposition of re-
sponses to white noise (Fournier et al., 2011, 2014) (Figure 2F).
Further improvements still can be gained from analysis of synap-
tic signals (Fournier et al., 2011) rather than spikes. Other gener-
alized decomposition schemes have been developed (e.g., linear
[Pillow et al., 2008] or nonlinear [NIM model in McFarland et al.,
2013] models accounting for the combined activity of simulta-
neously recorded units and their functional interactions), with
varying degrees of success (Figures 2D and 2E). These models
are useful in that they can provide a finer functional dissection
of afferent circuits to a neuron. On the other hand, they implicitly
describe a neuron’s response in terms of feedforward inputs,
although we know that most of a cortical neuron’s inputs derive
largely from local reverberating circuits. This obvious conundrum
is usually ignored.
Bottleneck Issues
We identify three major issues with the RF concept:
(1) Models of receptive fields fail to consider the synaptic na-
ture of their underlying processes and the impact of local
recurrent networks. Classical system theory provides
spike-based phenomenological models of ‘‘equivalent’’
feedforward circuits whose predictive value is limited to
the input statistics of the training set. As indicated above,
the exploration of RF nonlinearities has led to a variety of
models whose applicability is often limited to particular
stimuli, stimulus ranges, or cortical areas. Claims that
we understand 60% of the variance of visual neuron firing
in V1 (Carandini et al., 2005) is therefore misleading,
because current models do not explain spike output
causally, nor do they account for the complex conduc-
tance and voltage dynamics observed experimentally
(Baudot et al., 2013; Haider et al., 2010; Monier et al.,
2008). Adding higher-order correction terms may help,
but it increases model complexity without guarantying
success over all stimulus ranges. Because features in
natural images often combine multiple orientations and
spatial frequencies in one location, filtering the visual
input through many nonlinear (complex-like) components
(each for a specific feature) may contribute to the detec-
tion of high-order correlations of the visual scene (LeCun
et al., 2015). Combining detailed nonlinear statistical and
biophysical modeling is another alternative. Known non-
linearities in the transfer between retina and cortex (e.g.,
rectification of retinal and thalamic outputs) are generally
Figure 2. Receptive Field Models in Mammalian Primary Visual Cortex(A–C) Functional linear-nonlinear-Poisson (LNP) Models for V1 Neurons, and their characterization using spike-triggered analyses (adapted from Rust et al.,2005). (A) The standard simple cell model, based on a single space-time oriented filter. The stimulus (left arrow) is convolved with the filter, and the output ispassed through a half-wave rectifying and squaring nonlinearity. This signal determines the instantaneous rate of a Poisson spike generator. (B) The ‘‘energymodel’’ of a complex cell, based on a pair of space-time oriented filters with a quadrature (90�) phase relationship (Adelson and Bergen, 1985). Each filter isconvolved with the stimulus, and the responses are squared and summed. The resulting signal drives a Poisson spike generator. (C) Generalized LNP responsemodel. The cell is described by a set of n linear filters (L), which can be excitatory (E) or suppressive (S). The model response is computed by first convolving eachof the filters with the stimulus. An instantaneous nonlinearity (N) governs the combination of excitatory and suppressive signals that drives a Poisson spikegenerator (P).(D and E) Adapted from Dan Butts, with permission. (D) Nonlinear modeling of stimulus processing: the nonlinear input model (NIM) model assumes that eachdirect input to the cell is already a LN process. The resulting receptive field can be seen as a LNLN cascade selective to multiple features (McFarland et al., 2013).(E) Generalized contextual receptive field: the contextual input is represented as multiple nonlinear LFP bands, each reflecting local recurrent contribution due tolaminar-specific intra-columnar processing. Model performance is improved by taking into account trial-to-trial changes in the global cortical state reflected bythe LFP signals.(F) Parallel bank model of synaptic integration (adapted from Fournier et al., 2011, 2014). Left: decomposition method based on a Volterra expansion truncated tothe second-order diagonal terms and a PCA analysis for separating excitatory and inhibitory subunits. Right: example of an intracellular recording of a layer IVstellate cell, filled with biocytin with its morphological reconstruction (bottom). Note that this simple receptive field (see linear kernel, top row) expresses at thesubthreshold level (Vm) three excitatory cross-oriented subunits (see X-Y plots, middle column) and one inhibitory non-oriented subunit.
Neuron
Perspective
ignored, as Sompolinsky and Shapley (1997) pointed out.
Moreover, RFs are usually described one feature at
a time (orientation, direction, disparity .), whereas RF
optimization should take all features into account simulta-
neously—a tricky problem given their sometimes conflict-
ing impact (Figure 2F).
(2) Receptive fields are not invariant. RFs are eminently
adaptive, and the balance between their linear and
nonlinear components can change with the dimension-
ality or the statistical properties of an input (Fournier
et al., 2011; Yeh et al., 2009) (Figure 2F). This has
been shown using nonlinear kernel estimation in V1 neu-
rons for different white-noise stimulus conditions. For
example, Fournier et al. (2011) showed that both the spa-
tio-temporal extent of linear and nonlinear kernels and
their relative weights depended on the spatio-temporal
density of the stimulus. A possible interpretation is that
the simple-like component of a RF derives from feedfor-
ward drive, complex-like components result from recur-
rent lateral connections, and that a neuron’s best RF
description depends on a variable, stimulus-dependent
balance between these components. This view may be
consistent with several studies (Nauhaus et al., 2008; Po-
lat et al., 1998; Sceniak et al., 1999), showing that the
Neuron 88, October 7, 2015 ª2015 Elsevier Inc. 115
Neuron
Perspective
lateral propagation of activity between adjacent cortical
units decreases substantially when stimulus contrast in-
creases. If similar effects exist for stimulus spatial or tem-
poral density, the cell apparent ‘‘complex-ness’’ would
vary with the stimulus due to changes in the balance be-
tween its lateral and feedforward inputs.
(3) Existing RF models only explain relatively simple res-
ponse features and may not be applicable to multimodal
processing. In the visual system of awake monkeys, for
instance, the responses of some V1 andmany V2 neurons
to illusory contours match the orientation of their RF and
their responses to rigid contours (see section 1). Yet these
RFs generally correspond to filters applied to regions of
stimulus space that are too small to account for the
long-range dependencies needed for illusory contours.
Beyond primary areas, neurons responsive to complex
features (e.g., faces) are described using no longer phys-
ical (defined by the stimulus) but perceptual (defined by
the observer) attributes; what is a RF for such neurons?
More generally, the single-neuron approach seems inad-
equate to define the complex nonlinear transformation
applied by brain circuits to map physical reality into the
perceptual space. A new mathematical framework, one
that goes beyond the RF concept, seems necessary.
This framework will also need to include a general way of treat-
ing multisensory interactions. Neural activity in associative areas
often co-varies with signals from different sensory channels. For
example, in the medial superior temporal area of the monkey ex-
trastriate visual cortex, neurons may respond to both visual and
vestibular inputs to encode head motion (Angelaki et al., 2011).
Inmonkeyprimary auditory cortex, information about sound iden-
tity carried by individual neurons is increased when those sounds
are paired with images (Kayser et al., 2010). In the posterior pari-
etal cortex of rats trained in an auditory visual task, neurons may
respondboth tovisual andauditory inputs,butwithdifferentmulti-
modal covariation rules (Raposo et al., 2014). In short, developing
a general framework suitable for multisensory responses will
require the co-parameterization of multiple modalities. Given the
practical difficulty of doing so with single modalities, one can
appreciate the problems facing us with multimodal conditions.
Dynamic Assembly Codes for PerceptionThe variety of response properties found in a typical sensory area
of cortex suggests that the knowledge of the simultaneous state
of many cortical neurons should correlate better with descrip-
tions of perceptual representations. Measures of single-neuron
responses and receptive fields classically rely on firing rates.
But the timing of action potentials could also carry distributed in-
formation about a stimulus (Abeles, 1991). If spike timing is more
precise than the perceived temporal variations of a stimulus,
spike timing itself could be part of the code. In this case, informa-
tion would have to be decoded in a space defined by N neurons
of the population and P time-bins, where P is the ratio between
the shortest perceived stimulus variations and the uncompressi-
ble variability of spike timing.
Strong experimental evidence for such precise distributed
spatio-temporal patterns exists in the olfactory system of insects
116 Neuron 88, October 7, 2015 ª2015 Elsevier Inc.
(Laurent, 2002). In mammals, several forms of temporal codes
have been proposed. They include spike-latency codes (spike
timing relative to stimulus onset) (Spors et al., 2006; Van Rullen
and Thorpe, 2001) and spike-phase codes (e.g., relative to hip-
pocampal theta rhythm (Huxter et al., 2003). These paradigms
can be included in the more general distributed spike-sequence
codes proposed by Abeles (‘‘synfire chain’’ hypothesis [Abeles,
1991]). The temporal precision of stimulus-locked activity in
mammalian retina and thalamus neurons (2–5 ms) (Reinagel
and Reid, 2000) is at least compatible with such ideas. In themo-
tor cortex of behaving monkeys, temporal motifs have been
found to be correlated with the nature of a behavioral task (Vaa-
dia et al., 1995) and reinforcement expectation (Riehle et al.,
1997), but the functional significance of these observations
(made by the external observer and not forcibly decoded by
the neurons) is debated because of statistical issues (Ayzenshtat
et al., 2010) and because of incompatibility (a priori) with cortical
dynamics, generally unstable (London et al., 2010).
Temporal codeshavebeen studied in rodent olfaction, inwhich
odor sampling is paced by active sniffing; below the timescale of
sampling (the sniff period), temporal fluctuations of the input can
probably not be perceived, implying that response dynamics, if
they exist, do not encode stimulus fluctuations. In rodents,
odorant binding to odorant receptors (ORs) varies across the
ORs, leading to sequential activation of odorant sensory neurons
(Spors et al., 2006). These sequences likely contribute to the dy-
namic activation of the first relay neurons, themitral cells (MCs) of
the olfactory bulb (Bathellier et al., 2008). Odor identity can be
better decoded fromMCassemblies using temporal than rate co-
des (Cury and Uchida, 2010). Using optogenetic approaches,
Haddadet al. (2013) showed that downstreamneurons in piriform
cortex can use temporal information to adjust their firing rate but
that piriform neuron assemblies do not carry additional informa-
tion in their relative spike timing (Miura et al., 2012). Thus, the ex-
istence and use of temporal codes may be specific only to some
relay stations and pathways, possibly used for specific computa-
tions.Recent studies in auditory andsomatosensory cortex show
a good correlation (Bathellier et al., 2012), possibly even a causal
link (Musall et al., 2014; O’Connor et al., 2013), between cortical
rate codes and perception. But the simplicity of the stimuli and
the behavioral model used in these studies do not exclude the
possibility that cortical structures with a high information load
such as primate V1 use a sparse code based on precise spike
timing, as observed during high-dimensional natural-scene stim-
ulation (Baudot et al., 2013; Vinje and Gallant, 2000).
Whetheror notcertainneuronsaresensitive to thefine temporal
features of their input, a pragmatic approach is to assesswhether
population dynamics can be related to sensation and perception.
This approach has been used successfully in olfaction, in which
stimulus space cannot be parameterized,makingRF approaches
intractable. Population state analysis requires simultaneous re-
cordings from large and representative sets of neurons in an
area of interest (e.g., usingmultielectrode arrays or optical indica-
tors). It then aims at comparing the representations of stimuli us-
ing neural population metrics to infer the structure of sensory,
perceptual, or behavioral space. This approach skirts the com-
plex issue of deriving mathematical models of receptive fields to
fit the data, but developing it into a predictive tool remains a
0.250
1 20Trial #
Cel
l num
ber
87
9.5 kHz
1
26.8 kHz
Correlation
50 µm
25%
5 s
A B mixtures
population response switch
Figure 3. Switch-like Reconfigurations in Local Population Activity Patterns of Mouse Auditory Cortex during Auditory Perception(A) Top: 200 3 200 mm image of OGB1-stained neurons in the mouse auditory cortex. The line scan path used for two-photon calcium imaging in isoflurane-anaesthetized mice is superimposed. Bottom: typical calcium signals from all the neurons shown in (A) during presentation of various short sounds.(B) Single-trial population responses (left column) and similarity matrices (right column) for different mixtures of two pure tones (spectrograms on the top) thatexcited two distinct stereotypical response patterns. This figure shows that although the mixtures were varied gradually, the population response patternchanged in a switch-likemanner for an intermediatemixture ratio (see arrowhead; first response pattern elicitedwith mixtures dominated by 26.8 kHz, blue frame;second pattern, red). Adapted from Bathellier et al., 2012.
Neuron
Perspective
challenging task (Rabinovich et al., 2008). Population analysis has
beenextensivelyapplied to theolfactorybulbof fish (Friedrichand
Laurent, 2001), locust (Stopfer et al., 2003), and themouse (Bath-
ellier et al., 2008). It revealed that olfactory representationsdisplay
strong temporal dynamics even during constant stimuli, suggest-
ing that thedynamicsarepart of theolfactory representation.With
populationmetrics anddimensionality reduction techniques, sim-
ilarity between stimulus representations could be estimated, and
subspacescorresponding to representationsof thesameodorant
at different concentrations were identified (Stopfer et al., 2003),
providingmeans toaddress thedifficult question of concentration
invariance in olfactory perception.
Population analysis techniques have recently been applied to
rodent auditory cortex (Bathellier et al., 2012). Auditory cortex
studies contributed two important pieces of information. First, ac-
tivity in sensory cortex includes spontaneous but coordinated and
complex activity patterns that may reflect a repertoire of intrinsic
cortical dynamics with consequences on responses to sensory
input (Luczak et al., 2009). Importantly, sound-evoked cortical dy-
namics evolve in discrete steps when sounds are varied gradually
(Bathellier et al., 2012) (Figure 3). Altogether, these data suggest
that cortical circuits not only respond to their inputs but produce
their own dynamics in which sensory representations are
embedded, possibly resonating within specific ‘‘attractor’’ states.
Further investigation will be required to fully understand the inter-
play between intrinsic circuit dynamics and neural coding in cor-
tex, but an interestingworking hypothesis could be that the attrac-
torsofcorticaldynamics representelementary tokensofpercepts.
Bottleneck Issues
While increasing evidence points toward the existence of coordi-
nated dynamics in cortical circuits involving large cell assemblies
during perception, a difficult challenge still to overcome is to
prove their importance in perception.Development of large-scale
recording techniques and population analysis methods is one
important aspect, but as much as receptive field approaches,
which they generalize, they will detect only correlations between
brain activity and perception. This will be enough to infer new
algorithms potentially at play in sensory cortex. But to prove
that these algorithms are sufficient to generate perception, two
important tests should be performed. On one side, candidate
algorithms should be explicitly simulated to assess whether
they can quantitatively help reproducing perceptual capabilities
(e.g., object recognition) in a well-controlled theoretical setting,
thereby demonstrating that these algorithms perform the ex-
pected computations (in Marr’s sense). On the other side, to
show their causal involvement in perception in the living system,
the ideal experiment would be to show that, by replaying some
spatio-temporal activity pattern within or across brain areas,
one is able to trigger the emergence of the same percept actually
reported during observation of this very pattern. Such a Ge-
danken experiment, and quantitative tests on the impact pro-
duced by changing the firing rate or even jittering the individual
spike timingof cells composing the sameassembly,were already
proposed by Christoph von der Malsburg (von der Malsburg,
1986) andmay become possible in the near future. Cortical activ-
ity ‘‘re-encoding’’ experiments based on optogenetic tech-
niques, in which an illusory percept is produced by direct cortical
stimulation, recently started tobe tractable in rodents using opto-
genetics for very simple percepts such as object detection
(O’Connor et al., 2013) with the whisker system. So far, current
optogenetics is only able to precisely control the timing of spikes
on small or large populations of neurons, potentially identified by
genetic markers. However, creating precise spatio-temporal
sequences at a scale of the network that can be relevant for a
Neuron 88, October 7, 2015 ª2015 Elsevier Inc. 117
Neuron
Perspective
general percept is still a serious issuedespite the significant tech-
nical advances recently obtained using spatial light modulators
permitting the asynchronous activation of tens of individual cells
over a few hundred microns in awake mice (Szabo et al., 2014).
From Functional Principles to Neural Architectures forPerceptionBeyond the determination of algorithm principles (in David
Marr’s sense) thatmay underlie perception, another fundamental
question is their biological implementation. We here review a
few examples in which sensory processing principles could be
explained by particular cortical architectures and discuss the
generality of these architectures across senses and species.
As a first example, an important principle of visual processing is
‘‘spatial opponency,’’ which corresponds to suppressive effects
across features (contrast, motion, color contrast, orientation
contrast, center-surround) locatednearby in the retinotopic space
(e.g., center-surroundRF from retina to cortex). This effect can be
mathematically written as a convolution with a Mexican hat-
shaped filter equivalent to the opposite of the second-order
spatial derivative in visuotopic space (Ratliff, 1965). The role of
such operator is to emphasize local contrast and this type ofmotif
explains the Mach band illusion (Figure 1B) (Ratliff, 1965). It also
approximates the Gaussian Laplacian operator used in artificial
vision to extract edges and contours (Marr, 1982) and plays a
role in orientation selectivity. ‘‘Spatial opponency’’ can be imple-
mented in topographically organizedcircuitsby thespatially local-
ized lateral inhibitionarchitectures (Figure4A),whichareobserved
inmanysensory circuits includingprimary sensory neocortical cir-
cuits. Such architectures correspond to various forms of either
feedforward or recurrent inhibition (Figure 4B) when the interneu-
rons involved sendconnectionsmostly to spatially closecells (i.e.,
with neighboring and similar RF, e.g., Figures 4C and 4D). This
type of lateral inhibition is found in many different sensory struc-
tures, such as the retina or primary visual cortex in mammals,
with sometimes further refinements. For example, in cat, ferret,
and macaque, orientation selectivity in simple cells in the tha-
lamo-recipient layer 4 (4Cb for the macaque) is known to result
froma push-pull circuit in which the excitatory neurons receive in-
hibition from similarly orientation-tuned interneurons but in oppo-
sition of phase (Figures 4E and 4F) (Troyer et al., 1998).
While local connectivity motifs such as lateral inhibition under-
lie the extraction of precise local relationship in a visual scene
and probably explain the core of receptive fields in V1, it is un-
likely that they account for the emergence of global percepts
based on inputs spanning large retinotopic distances. However,
the later may in part result from long-distance connectivity motifs
in the cortical circuit, so as to bind together elemental visual fea-
tures in a meaningful way (Figure 5A). In higher mammals (tree
shrew, ferret, and cat as well as non-human primates) recon-
structed pyramidal cell axons that remain within the gray matter
extend over several hypercolumns (up to 6–8 mm in the tree
shrew [Bosking et al., 1997]; in the cat [Gilbert and Li, 2012],
but seeMartin et al. [2014]). Supposedly, these long connections
modulate the response gain of the local circuits (i.e., hypercol-
umn), inducing often suppressive effects although particular
center-surround stimulus conditions can induce specific boost-
ing (Sillito et al., 1995).
118 Neuron 88, October 7, 2015 ª2015 Elsevier Inc.
In spite of this uncertain status, horizontal connectivity has
long been presented as the biological substrate of iso-prefer-
ence binding. This functional organization principle is derived
from a developmental canonical rule which posits that ‘‘who fires
together (or is alike) tend to wire together’’ and has been imple-
mented successfully by LISSOM models to account for the
development of lateral connectivity in sensory cortical maps
(Miikkulainen et al., 2006). At the psychophysical level, this
view corresponds to the perceptual ‘‘association field’’
(Figure 5B) (Field et al., 1993). This concept assumes the facilita-
tion of collinear and, to a lesser extent, co-circular spatial inte-
gration of oriented contrast edges. This elegant psychophysical
hypothesis accounts in humans for the ‘‘pop-out’’ perception of
smooth contiguous path integration even when immersed in a
sea of randomly oriented edge elements (Field et al., 1993)
(Figure 5A) and the facilitation of target detection by high-
contrast co-aligned flankers (Polat and Sagi, 1993) (Figure 5A).
At the neuronal level, this view is supported by the electrophys-
iological demonstration of a ‘‘neural facilitation field’’ (Gilbert and
Li, 2012) corresponding to a boosting of the response gain to an
optimally oriented contrast edge within the classical RF when
flankers were simultaneously flashed in the immediate ‘‘silent
surround’’ and co-aligned along the preferred orientation axis
of the recorded cell (Figures 5C and 5D). Remarkably, this
gain-control effect seems to depend on top-down signals, as it
is weakened by diverted attention and suppressed by anes-
thesia (Li et al., 2006).
Further evidence and understanding comes from intracellular
data and VSD imaging in the anesthetized cat, demonstrating
long-distance propagation of visually evoked subthreshold ac-
tivity through lateral (and possibly feedback) connectivity outside
the classical receptive field (intracellular: Bringuier et al., 1999;
Fregnac, 2012; VSD: Benucci et al., 2007). More recently, both
techniques have been combined to show that a critical threshold
of spatial synergy and temporal summation has to be crossed
before the impact of the long-range interactions (in themV range)
can be functionally detected (Chavane et al., 2011). Taken
together, these studies suggest the existence (in higher mam-
mals) of structuro-functional collinearity biases, intrinsic to V1
and present at a subthreshold level in the anesthetized brain,
which require attention and feedback from higher cortical areas
to be expressed at a spiking level. No evidence for such circuits
has been found in the rodent.
Bottleneck Issues
One difficulty in identifying canonical motifs is that the functional
expression of cortical circuits adapts continuously to the statis-
tics of the sensory environment. Moreover, most associative
plasticity algorithms have been introduced to build distributed
memories of our environment, during development or learning
(review in Fregnac, 2003). Their implication in perception itself,
although envisioned by James and Hebb, remains largely unex-
plored. The dominant plasticity algorithms are derived from cor-
relation-based rules (Bienenstock et al., 1982), which extend
beyond Hebb’s principle, and are often seen as a universal set
of recipes to build long-lasting assemblies or reinforce causal
chains in processing. Although earlier work focused on syn-
chrony (with a ± 50 ms temporal contiguity window (e.g., Baranyi
and Feher, 1981), the refinement of in vitro techniques showed
Figure 4. Canonical Circuits(A) Lateral inhibition (illustration of center-surround interactions, taken from Cavanaugh et al., 2002). The ratio of Gaussians (RoG) model is constructed fromindependent and spatially stable center and surround components with two Gaussian envelopes of different spatial spread (s).(B) Major types (adapted from Isaacson and Scanziani, 2011) of intracortical inhibition: left, feedback; right, feedforward. Inhibitory cell in blue (round soma);excitatory cell in gray (pyramidal soma). Red axon emitted by the postsynaptic target cell (left, feedback) or by a common excitatory drive (right, feedforward).(C) Sensory-evoked feedforward inhibition (FFI): inhibition enforces precise spike timing in a principal cell in auditory cortex (adapted fromWehr and Zador, 2005):top, timing of action potentials; middle, subthreshold membrane potential; bottom, underlying excitatory and inhibitory synaptic conductances. Action potentialslargely occur only in the narrow time window during which excitation precedes inhibition.(D) Visual cortex of the rat (Liu et al., 2010). Stimulus selectivity in the rodent cortex emerges from a temporal shift in the timing of excitation relative to inhibition.(E) Push-pull in cat, ferret, and higher mammals (adapted from Troyer et al., 1998 and Jens Kremkow, with permission): prototypic recurrent network model oflayer 4 in the mammalian visual cortex ‘‘V1’’ with correlation-based connectivity implementing the push-pull receptive field organization. Inputs from the LGNprovide direct excitatory (push). In cat V1, inhibitory neurons project preferentially to neurons having a receptive field phase difference of around 180�, effectivelyimplementing the pull inhibition. Note also the intracortical reciprocal inhibition between inhibitory I1 and I2 neurons and the intracortical excitatory amplificationfor E1 and E2 neurons.(F) Generalized forms of push-pull architectures: the push-pull organization demonstrated for spatial phase can be hypothetically generalized to other dimensionsthan space such as orientation and direction (Monier et al., 2003, 2008).
Neuron
Perspective
that the temporal order of pre- and postsynaptic spikes is critical
for the sign of the synaptic change. The experimentally defined
Spike-Timing-Dependent Plasticity (STDP+) rule that is used in
most computational models is a modified Hebbian rule
(Figure 6A) that includes a potentiation part when the postsyn-
aptic spike follows shortly the presynaptic spike (necessary for
strengthening causal links) and a depression part for the reverse
order association (necessary for network stability) (Bi and Poo,
1998; Markram et al., 1997). However, its ubiquity remains
debated since other forms of STDP have also been observed,
where the ‘‘causal’’ pre/ post spike sequence induces depres-
sion and the reversed sequence (post / pre, or ‘‘anti-causal’’)
induces potentiation (Bell et al., 1997; Fino and Venance, 2011;
Letzkus et al., 2006; Safo and Regehr, 2005) (Figure 6A). There
also exist synapses with symmetric STDP between layer IV spiny
stellate cells in layer 4 of S1 cortex (Egger et al., 1999).
Interestingly, similar STDP rules have been proposed by some
daring theoreticians (Von der Malsburg, 1981) to operate also on
fast timescales (tens ofmilliseconds), compatible with the waxing
andwaning of a percept. The fast reversible formof STDP that re-
lates the best to perception is anti-Hebbian (STDP�). It has been
described in vivo and in vitro in the electrosensory lobe (ELL) of
the mormyrid electric fish (Bell et al., 1997) and should not be
confounded with classical LTD, in view of the fast time constant
and the self-erasing feature of negative STDP (Figure 6A).
The originality of the STDP� rule as a decorrelation algorithm
reducing input redundancy was theoretically envisioned first in
the context of the visual system (Barlow and Foldiak, 1989). But
it was first demonstrated experimentally in the ELL of the fish
(Bell et al., 1997) in neurons that compare the current sensory
input (electric field image encoded by electrosensory afferents
to the ELL) and the efference copy signal (originating in the electric
Neuron 88, October 7, 2015 ª2015 Elsevier Inc. 119
Figure 5. The ‘‘Perceptual Association Field’’(A) Top: ‘‘pop-out’’ emergence of a continuous integration path in a sea of randomly oriented Gabor patches (Field et al., 1993). Bottom: facilitation of detection ofa low-contrast vertical Gabor element induced by the simultaneous presentation of co-aligned high contrast flanker elements (Polat and Sagi, 1993).(B) Hypothetical association field induced by an oriented element through lateral interactions (Field et al., 1993).(C) The ‘‘iso-functionalbinding’’ hypothesis (adapted fromGilbert andLi, 2012). An individual superficial layer cortical pyramidal cell forms long-rangeconnections thatextend many millimeters parallel to the cortical surface. Long-range connections (> 500 mm from the injection center) tend to link columns of similar orientationpreference.(D) The ‘‘neural facilitation field’’ (Gilbert and Li, 2012). Left: the responses of V1 neurons are amplified in the awake behaving monkey by collinear contoursextending outside the RF. Introducing a cross-oriented bar between the collinear segments blocks the contour-related facilitation. Right: two-dimensional map offacilitatory (blue) and inhibitory (red) modulation of the response to an optimally oriented line segment centered in the RF (horizontal white bar). The spikingmodulation is suppressed by anesthesia.
Neuron
Perspective
fish from premotor neurons of the electric organ). The efference
copy updates the integration of the new sensory input by sub-
tracting the previous sensory reafference. The resulting functional
impact is to depress the transmission of inputs whose occurrence
reflects past evoked discharge and to strengthen unexpected in-
formation arising from the environment (Bell, 1981). The general-
ization of this fast-acting adaptive rule to other sensory systems
during sensory exploration suggests a plausible model for the
cortical the visuo-oculomotor system of higher mammals. One
can envision (as developed in Figures 6B and 6C) that an effer-
ence copy generated by saccadic oculomotor activity planning
in higher mammals (Crapse and Sommer, 2008) filters out in the
early visual system the predictable retinal changes due to
voluntary movements. This would imply that synapses onto thal-
amus and perigeniculate cells, which convey the contextual
120 Neuron 88, October 7, 2015 ª2015 Elsevier Inc.
feedback or prediction from the cortex, obey STDP� rules acting
on the timescale of a percept (100 ms), which remains to be
demonstrated. Since re-entrant cortico-thalamic loops are
shared by many sensory systems (review in Briggs and Usrey,
2008) this idea could extend to other sensory modalities.
Together, the examples of connectivity schemes and possible
diversity of fast plasticity algorithms presented above show the
difficulty of assigning particular circuit motifs to every receptive
field properties or operators (e.g., feedforward ON-OFF kernels
[Vidyasagar and Eysel, 2015], lateral diffusion kernels, or cor-
tico-thalamic feedback kernels) or even to the non-stationarities
and context dependencies of receptive fields asmentioned earlier
(long-range lateral connections). The same circuit can have facil-
itatory or depressive effect depending on the considered time-
scale and sensory context. What makes it difficult to extend our
A
C
B Figure 6. Plasticity Algorithms andPerceptual Filtering(A) Polymorphy of associative plasticity algorithms(adapted from Fregnac, 2003). The left panel de-tails different forms of spike timing-dependentplasticity rules established in vitro (co-cultures andacute slices). The induced synaptic change is ex-pressed as a function of the temporal delayseparating postsynaptic firing from presynapticfiring (taken here as the synchrony reference),imposed during the pre-post pairing protocol.From top to bottom: pyramidal cells in hippo-campus or in non-granular layers in neocortex,granular spiny stellate cells in neocortex,GABAergic neurons in hippocampal cultures,GABAergic medium ganglionic layer cells in theelectrosensory lobe (ELL) of the mormyrid electricfish. Synaptic potentiation in pink, depression inblue.(B and C) Prediction of negative STDP in feedbackprojections from higher areas in the early visualsystem. (B) Perceptual filtering in the ELL of themormyrid electric fish at the level of the largePurkinje cells of the ELL, by the efferent copyoriginating from the torus circularis nucleus, pre-motor to the electrical discharge center (EOD). (C)Hypothetical predictive filtering in the early visualsystem. Negative STDP, similar to that describedin the ELL, is expected synapses that project thecortical context feedback back to the sensorythalamic gate. The cortical prediction is comparedto what is received from the retina, resulting in thecancellation of expected correlations.
Neuron
Perspective
understanding of the circuit bases of perception is, first, that the
algorithmic principles of fast reversible plasticity are not fully eluci-
dated (see above), and second, that a large diversity of circuit
motif and mechanisms exists across senses and species.
We have seen that certain circuit principles (e.g., lateral inhibi-
tion) depend on the precise, topographic organization of the en-
coded features. Examples of topographically organized cortical
modules exist in various species and sensory modalities: the
whisker barrels in rodents, the visual hypercolumns in higher
mammals, as well as the blob and interblob patches in the visual
cortex of primates. Although their role is not fully elucidated (Hor-
tonandAdams, 2005), the existenceof thesemodulesmay reflect
the need to organize large flows of information in away that eases
the application of relational rules between elements of the unisen-
sory scene and extract invariant representations (Kavukcuoglu
et al., 2009). To take one concrete example, rodents that are often
nocturnal or extensively travel in underground tunnels (like rats or
mice) require well-structured tactile perception through their
whisker pads. However, what is required for haptic perception
does not seem to be needed for vision in rodents: accordingly,
even if neuronsof thevisual cortexof the rodent exhibit orientation
tuning, they are not spatially clustered according to their orienta-
tion preference (Ohki et al., 2006). This ‘‘salt and pepper’’ organi-
zationof V1 in the rodent contrastswith thecontinuousorientation
preference domains observed in diurnal predators such as cats
andnon-humanprimates. Thisdifference in architecturemaypre-
serve some general principles of network organization between
rodents and more visual mammals with orientation preference
maps. For example, increased connectivity probability between
cells with similar feature selectivity exist both in mice (Cossell
et al., 2015) and in cats or ferrets on the scale of an hypercolumn
(but see above for species-specific constraints for longer-dis-
tance connectivity). But inmice, due to the salt-and-pepper orga-
nization, feedforward inhibition colocalized with excitation domi-
nates in contrast to the spatial segregation of excitatory and
inhibitory sources mediating the push-pull organization of tha-
lamo-recipient RF in layer 4 in the cat and ferret and layer 4b in
monkey V1 (Figures 5D–5F). In mice, interneurons receive inputs
from cells with different orientation tuning (Bock et al., 2011) and
therefore control the global response gain rather than the sharp-
ness of orientation selectivity (Atallah et al., 2012), which is poor
in mouse cortex and mostly arises from thalamic inputs (Lien
and Scanziani, 2013). This interspecies shift in microcircuit archi-
tecture,where functional selectivitymaybe already established in
thalamus for rodents, while requiring cortex for highermammal, is
reflected at the computational level, considering for example the
much higher orientation detection performance in cats and pri-
mates than in mice (Andermann et al., 2010).
Along the same line, long-distance projections of cortical
neurons may also serve very different functions in rodents
compared to higher mammals. Recent connectomics studies
show that, in the rodent cortex, axons fromprimary sensory areas
can massively reach motor or decisional areas and vice versa
(Zingg et al., 2014). In ferret, cat, and tree shrew, long-range
Neuron 88, October 7, 2015 ª2015 Elsevier Inc. 121
Neuron
Perspective
connections prevalently intrinsic to the same primary area, e.g.,
V1 for vision, are superseded by a stronger involvement of cor-
tico-cortical feedback in non-human primates (Angelucci et al.,
2002) and in humans. In primates, the reach of long-distance
axons is limited to the representation of only a few degrees in
foveal vision, and cortico-cortical loops may be required to go
beyond this spatial diffusion boundary. The selective pressure
produced by the refined grain of the retinotopic map in primates
(resulting in an increase in retino-cortical magnification factor)
and the relative size increase of cortical areas may render very
long-range intrinsic horizontal axons ineffective (possibly too
slow, toodiluted, or tooexpensive) tomediate long-distance cen-
ter-surround interactions in the visual field. Therefore, these
lateral interactions are established in primates via higher areas
at the expense of intra-V1 connectivity.Finally, interspecies differences also exist for the circuits pro-
cessing multisensory information. In primates, very few neurons
project directly from one primary sensory area to another (Falch-
ier et al., 2002), and multimodal responses remain subthreshold
except during precise spatial or temporal coincidence (Stein and
Stanford, 2008). In rodents, cross-modal connectivity is strong
across primary sensory areas (Laramee and Boire, 2014) and
leads to suprathreshold responses (Olcese et al., 2013). Hence,
the classical view of segregated feedforward streams of unimo-
dal sensory information before associative areas is mostly appli-
cable to higher vertebrates and not so much to the rodent. In the
rodent, multimodal interactions can reshape unisensory repre-
sentations already in primary cortical areas, potentially building
cross-modal associations with a lower level of complexity than
associations between sensory modalities made by primates.
These examples together show that it is unavoidable to
choose a comparative approach in the search for the archetypal
circuits of perception and discriminate, based on structural and
ecological evidence, the common principles from the circuit mo-
tifs that reflect a particular use of a sensory modality.
ConclusionThis overview illustrates the complexity of explaining perceptual
processes in terms of realistic neural-based architecture and
rules. A bottom-up strategy on its own seems doomed to fail;
conceptual approaches must be developed to reduce structural
complexity (see Box 2).
Current observations and models suggest that the psycho-
logical laws of visual ‘‘pop-out’’ non-attentive perception rely
principally on the internal architecture of primary sensory areas
expressed through recurrent and horizontal connectivity and
probably shaped by correlation-based associative synaptic
plasticity. This architecture would enable the self-organization
and propagation of activity that ultimately facilitate plausible
feature associations. The rewiring of horizontal connectivity in
auditory cortex of the developing ferret induced by the forced re-
innervation by visual afferents (Sharma et al., 2000) after lesion of
the normal subcortical afferent pathway, and turning it function-
ally in a visual cortex, is today the most impressive demonstra-
tion of the versatility of sensory cortical modules.
Experimentally, the technical possibility of recording simulta-
neously large assemblies of functionally identified neurons with
millisecond precision opens a new era. On the theoretical front,
122 Neuron 88, October 7, 2015 ª2015 Elsevier Inc.
many paths exist toward modeling the sensory brain, even in
cases as seemingly simple as unimodal, low-level perception
and form recognition in the absence of behavioral attention.
Others argue for ‘‘top-down’’ approaches, via the abstract math-
ematical modeling of network-scale cortico-cortical and cortico-
thalamo-cortical loop architectures (Bastos et al., 2012; Mum-
ford, 1991). They often describe the brain as aBayesian inference
device. This approachworks well to describe the computation (in
Marr’s sense), but it fails to identify algorithms and underlying
circuits. What is missing is a ‘‘middle-out’’ approach that can
identify plausible models to link architecture and cognition.
Large-scale brain initiatives (The Allen Institute, The Human
Brain Project) are currently developing data collection pipelines
and computational infrastructures mostly driven by a bottom-
up approach. This conceptual ‘‘prior’’ takes its ground in the
‘‘cortical column’’ ideation simulated in the Blue Brain Project
(Markram, 2012). It runs on the assumption that a detailed ‘‘bot-
tom-up’’ reconstruction and exhaustive simulation of neuronal
elements will eventually reveal canonical microcircuits, from
which the laws of perception will emerge. In this latter formula-
tion, these circuits are seen as innate building blocks of knowl-
edge for perception that are combined during learning to form
complex Lego constructs composed in such a way as to engram
new memories (Markram and Perin, 2011). The merit of this bold
approach in the rodent is that paradigms developed with pri-
mates (see the inspiring work of Bill Newsome and colleagues
that linked neurometric and psychometric perceptual measures
[Britten et al., 1992]) are now being applied to rodents with
measurable success (Bathellier et al., 2012; Musall et al.,
2014). But to avoid the trap of nested complexity, it remains
necessary to extract generic principles (Marr’s algorithmic level)
and obtain modeling simplifications. In this vein, one should be
careful not to force homologies (between distant species) that
may not always exist, and one should take advantage of the di-
versity of experimental systems, many of which have been
driving the field: the cat visual and somatosensory cortex for vi-
sual perception and early associative memory storage (Spinelli
and Jensen, 1979), the rat hippocampus and entorhinal cortex
for navigation-related computation (Moser and Moser, 2013;
O’Keefe and Recce, 1993), the electric fish electrosensory lobe
for perceptual filtering and novelty detection (Bell, 1981), the in-
sect olfactory system for associative learning of dynamic sen-
sory representations (Cassenaer and Laurent, 2012). We remain
convinced of the importance of ‘‘simpler’’ organisms to discover
canonical principles. ‘‘New directions in Science are launched by
new tools much more often than by new concepts’’ (Dyson,
1997). The development of optogenetics (mostly in rodents)
and its undeniable usefulness has triggered an historical change
in strategy in the study of sensory perception. The pressure to
standardize and the attraction of ‘‘big data’’ have led, too quickly
in our view, to a loss of perspective on diversity and evolution.
In this regard, the choice of the mouse model to study
mammalian vision seems odd to us. For example, the otherwise
laudable goal to industrialize and standardize data collection
could lead to a great impoverishment of visual protocols if one
is not careful enough. Recent studies show that themouse visual
system seems to depend heavily on locomotion (Niell and
Stryker, 2008), and the generalized use of ‘‘running-on-a-ball’’
Neuron
Perspective
paradigms have set a new sensory-motor behavior standard, far
from natural exploration behavior (Wallace et al., 2013). By repli-
cating pioneering experiments showing that active exploration is
needed to induce visually guided reaching behavior in kittens
deprived of vision (Hein and Held, 1967), Stryker and colleagues
identified in the rodent a specific ascending circuit, driven by
locomotion, and controlling sensory-induced plasticity. Howev-
er, if the whole-body locomotion seems needed in rodents, eye
movements and their proprioceptive inflow suffice in higher
mammals (Buisseret et al., 1978). In general, no report of a tight
relation between locomotion and vision has been confirmed in
cats, ferrets, and non-human primates in which primary sensory
and motor cortices are separated by numerous secondary and
associative areas (Markov et al., 2013). Hence, mouse vision
might acutely illustrate the species specificity of sensory-guided
behavior (see section 6).
Comparative studies show that primate brains are not simply
inflated versions of rodent brains. As underlined by Anthony
Movshon (Movshon, 2013), typical long-distance axons in ro-
dents do not only remain within their cortical area of origin as
in the ferret or the cat and rather tend to link multiple areas, sen-
sory, limbic, and motor. If long-range connections underlie our
‘‘perceptual grammar,’’ mice may actually have a very different
perceptual language than higher mammals.
But if, as we argue, the choice of rodents as the reference
model may be a deceiving alley for studying visual processing,
it may come at its advantage when searching for mechanisms
responsible for olfaction, tactile sensing, or multimodal integra-
tion. The reduced size of the computational sheet makes that
the higher visual cortical areas of the mouse abut directly other
primary areas such as S1 and A1, with their interfacing border
constituting an ideal site for multimodal integration (Laramee
and Boire, 2014). This strong connectivity results in suprathres-
hold multimodal interaction even in mouse primary sensory
cortex (Olcese et al., 2013), whereas such influences are sub-
threshold in primates (Stein and Stanford, 2008). In that respect,
the mouse may offer interesting opportunities to understand the
functional significance of heteromodal influences in primary
areas, otherwise present but silent in primates.
Hence, future progress in data-driven modeling of unimodal
and multimodal perception will depend on a judicious choice
of experimental models, on the simultaneous acquisition of neu-
ral data at many spatial and temporal scales, and on the defini-
tion of naturalistic sensory input benchmarks that meaningfully
constrain quantitative data-driven simulations. These goals
emphasize the inescapable roles of well-designed experiments
and of comparative approaches.
ACKNOWLEDGMENTS
We would like to thank Gilles Laurent, whose expert help has helped us toimprove the clarity and focus of this review. We would like to thank also KirstyGrant for her expert views on anti-Hebbian plasticity, and our colleagues CyrilMonier, Daniel Shulz, and Florian Gerard-Mercier for helpful comments. Wealso would like to extend our thanks to the participants of the NYU workshopon canonical plasticity held at La Pietra Villa in July 2015, where some of theseviews were discussed. We acknowledge support from CNRS, The Paris-Sa-clay IDEX (NeuroSaclay and I-Code), the French National Research Agency(ANR: NatStats Sensemaker and Complex-V1), the European Community(FET-Bio-I3 integrated programs IP FP6: FACETS [015879], IP FP7: BRAIN-
SCALES [269921]; Marie Curie CIG [334581]; FET-Open Brain-i-nets[243914]; FET-Flagship: The Human Brain Project [604102]), The InternationalHuman Frontier Science Program Organization (CDA-0064-2015), and theFyssen foundation.
REFERENCES
Abboud, S., Maidenbaum, S., Dehaene, S., and Amedi, A. (2015). A number-form area in the blind. Nat. Commun 6, 6026.
Abeles, M. (1991). Corticonics neural circuits of the cerebral cortex (Cam-bridge University Press).
Adelson, E.H., and Bergen, J.R. (1985). Spatiotemporal energy models for theperception of motion. J. Opt. Soc. Am. A 2, 284–299.
Alonso, J.M., and Swadlow, H.A. (2005). Thalamocortical specificity and thesynthesis of sensory cortical receptive fields. J. Neurophysiol. 94, 26–32.
Andermann, M.L., Kerlin, A.M., and Reid, R.C. (2010). Chronic cellular imagingof mouse visual cortex during operant behavior and passive viewing. Front.Cell. Neurosci. 4, 3.
Angelaki, D.E., Gu, Y., and Deangelis, G.C. (2011). Visual and vestibular cueintegration for heading perception in extrastriate visual cortex. J. Physiol.589, 825–833.
Angelucci, A., Levitt, J.B., and Lund, J.S. (2002). Anatomical origins of the clas-sical receptive field and modulatory surround field of single neurons in ma-caque visual cortical area V1. Prog. Brain Res. 136, 373–388.
Atallah, B.V., Bruns, W., Carandini, M., and Scanziani, M. (2012). Parvalbumin-expressing interneurons linearly transform cortical responses to visual stimuli.Neuron 73, 159–170.
Ayzenshtat, I., Meirovithz, E., Edelman, H., Werner-Reiss, U., Bienenstock, E.,Abeles, M., and Slovin, H. (2010). Precise spatiotemporal patterns among vi-sual cortical areas and their relation to visual stimulus processing.J. Neurosci. 30, 11232–11245.
Baranyi, A., and Feher, O. (1981). Synaptic facilitation requires paired activa-tion of convergent pathways in the neocortex. Nature 290, 413–415.
Barlow, H.B. (1972). Single units and sensation: a neuron doctrine for percep-tual psychology? Perception 1, 371–394.
Barlow, H.B., and Foldiak, P. (1989). Adaptationa and decorrelation in the cor-tex. In The Computing Neuron, R. Durnin, C. Miall, and G. Mitchison, eds. (Ad-dison-Wesley), pp. 54–72.
Bastos, A.M., Usrey, W.M., Adams, R.A., Mangun, G.R., Fries, P., and Friston,K.J. (2012). Canonical microcircuits for predictive coding. Neuron 76,695–711.
Bathellier, B., Buhl, D.L., Accolla, R., and Carleton, A. (2008). Dynamicensemble odor coding in the mammalian olfactory bulb: sensory informationat different timescales. Neuron 57, 586–598.
Bathellier, B., Ushakova, L., and Rumpel, S. (2012). Discrete neocortical dy-namics predict behavioral categorization of sounds. Neuron 76, 435–449.
Baudot, P., Levy, M., Marre, O., Monier, C., Pananceau, M., and Fregnac, Y.(2013). Animation of natural scene by virtual eye-movements evokes high pre-cision and low noise in V1 neurons. Front. Neural Circuits 7, 206.
Bell, C.C. (1981). An efference copy which is modified by reafferent input. Sci-ence 214, 450–453.
Bell, C.C., Han, V.Z., Sugawara, Y., and Grant, K. (1997). Synaptic plasticity ina cerebellum-like structure depends on temporal order. Nature 387, 278–281.
Benucci, A., Frazor, R.A., and Carandini, M. (2007). Standing waves and trav-eling waves distinguish two circuits in visual cortex. Neuron 55, 103–117.
Bi, G.Q., and Poo, M.M. (1998). Synaptic modifications in cultured hippocam-pal neurons: dependence on spike timing, synaptic strength, and postsynapticcell type. J. Neurosci. 18, 10464–10472.
Neuron 88, October 7, 2015 ª2015 Elsevier Inc. 123
Neuron
Perspective
Bienenstock, E.L., Cooper, L.N., andMunro, P.W. (1982). Theory for the devel-opment of neuron selectivity: orientation specificity and binocular interaction invisual cortex. J. Neurosci. 2, 32–48.
Blanke, O. (2012). Multisensory brain mechanisms of bodily self-conscious-ness. Nat. Rev. Neurosci. 13, 556–571.
Blanke, O., Slater, M., and Serino, A. (2015). Behavioral, Neural, and Computa-tional Principles of Bodily Self-Consciousness. Neuron 88, this issue, 145–166.
Bock, D.D., Lee, W.C., Kerlin, A.M., Andermann, M.L., Hood, G., Wetzel, A.W.,Yurgenson, S., Soucy, E.R., Kim, H.S., and Reid, R.C. (2011). Network anat-omy and in vivo physiology of visual cortical neurons. Nature 471, 177–182.
Bonath, B., Noesselt, T., Krauel, K., Tyll, S., Tempelmann, C., and Hillyard, S.A.(2014). Audio-visual synchronymodulates the ventriloquist illusion and its neu-ral/spatial representation in the auditory cortex. Neuroimage 98, 425–434.
Bosking, W.H., Zhang, Y., Schofield, B., and Fitzpatrick, D. (1997). Orientationselectivity and the arrangement of horizontal connections in tree shrew striatecortex. J. Neurosci. 17, 2112–2127.
Botvinick, M., and Cohen, J. (1998). Rubber hands ‘feel’ touch that eyes see.Nature 391, 756.
Briggs, F., and Usrey, W.M. (2008). Emerging views of corticothalamic func-tion. Curr. Opin. Neurobiol. 18, 403–407.
Bringuier, V., Chavane, F., Glaeser, L., and Fregnac, Y. (1999). Horizontal prop-agation of visual activity in the synaptic integration field of area 17 neurons.Science 283, 695–699.
Britten, K.H., Shadlen, M.N., Newsome, W.T., and Movshon, J.A. (1992). Theanalysis of visual motion: a comparison of neuronal and psychophysical per-formance. J. Neurosci. 12, 4745–4765.
Buisseret, P., Gary-Bobo, E., and Imbert, M. (1978). Ocular motility and recov-ery of orientational properties of visual cortical neurones in dark-reared kittens.Nature 272, 816–817.
Carandini, M., and Heeger, D.J. (2012). Normalization as a canonical neuralcomputation. Nat. Rev. Neurosci. 13, 51–62.
Carandini, M., Demb, J.B., Mante, V., Tolhurst, D.J., Dan, Y., Olshausen, B.A.,Gallant, J.L., and Rust, N.C. (2005). Do we know what the early visual systemdoes? J. Neurosci. 25, 10577–10597.
Cassenaer, S., and Laurent, G. (2012). Conditional modulation of spike-timing-dependent plasticity for olfactory learning. Nature 482, 47–52.
Cavanaugh, J.R., Bair, W., andMovshon, J.A. (2002). Nature and interaction ofsignals from the receptive field center and surround in macaque V1 neurons.J. Neurophysiol. 88, 2530–2546.
Chavane, F., Sharon, D., Jancke, D., Marre, O., Fregnac, Y., and Grinvald, A.(2011). Lateral Spread of Orientation Selectivity in V1 is Controlled by Intracort-ical Cooperativity. Front. Syst. Neurosci 5, 4.
Chen, L.M., Friedman, R.M., and Roe, A.W. (2003). Optical imaging of a tactileillusion in area 3b of the primary somatosensory cortex. Science 302, 881–885.
Cossell, L., Iacaruso, M.F., Muir, D.R., Houlton, R., Sader, E.N., Ko, H., Hofer,S.B., and Mrsic-Flogel, T.D. (2015). Functional organization of excitatory syn-aptic strength in primary visual cortex. Nature 518, 399–403.
Crapse, T.B., and Sommer, M.A. (2008). Corollary discharge circuits in the pri-mate brain. Curr. Opin. Neurobiol. 18, 552–557.
Cury, K.M., and Uchida, N. (2010). Robust odor coding via inhalation-coupledtransient activity in the mammalian olfactory bulb. Neuron 68, 570–585.
David, S.V., Vinje, W.E., and Gallant, J.L. (2004). Natural stimulus statisticsalter the receptive field structure of v1 neurons. J. Neurosci. 24, 6991–7006.
De Weerd, P., Desimone, R., and Ungerleider, L.G. (1996). Cue-dependentdeficits in grating orientation discrimination after V4 lesions in macaques.Vis. Neurosci. 13, 529–538.
Diaz-Caneja, E. (1928). Sur l’alternance binoculaire. Annales d’oculistique(Paris) 165, 721–731.
124 Neuron 88, October 7, 2015 ª2015 Elsevier Inc.
DiCarlo, J.J., Zoccolan, D., and Rust, N.C. (2012). How does the brain solvevisual object recognition? Neuron 73, 415–434.
Dyson, F. (1997). Imagined Worlds (Harvard University Press).
Eagleman, D.M. (2001). Visual illusions and neurobiology. Nat. Rev. Neurosci.2, 920–926.
Egger, V., Feldmeyer, D., and Sakmann, B. (1999). Coincidence detection andchanges of synaptic efficacy in spiny stellate neurons in rat barrel cortex. Nat.Neurosci. 2, 1098–1105.
Estebanez, L., El Boustani, S., Destexhe, A., and Shulz, D.E. (2012). Correlatedinput reveals coexisting coding schemes in a sensory cortex. Nat. Neurosci.15, 1691–1699.
Falchier, A., Clavagnier, S., Barone, P., and Kennedy, H. (2002). Anatomicalevidence of multimodal integration in primate striate cortex. J. Neurosci. 22,5749–5759.
Field, D.J., Hayes, A., and Hess, R.F. (1993). Contour integration by the humanvisual system: evidence for a local ‘‘association field’’. Vision Res. 33, 173–193.
Fino, E., and Venance, L. (2011). Spike-timing dependent plasticity in striatalinterneurons. Neuropharmacology 60, 780–788.
Fournier, J., Monier, C., Pananceau, M., and Fregnac, Y. (2011). Adaptation ofthe simple or complex nature of V1 receptive fields to visual statistics. Nat.Neurosci. 14, 1053–1060.
Fournier, J., Monier, C., Levy, M., Marre, O., Sari, K., Kisvarday, Z.F., and Fre-gnac, Y. (2014). Hidden complexity of synaptic receptive fields in cat V1.J. Neurosci. 34, 5515–5528.
Fregnac, Y. (2003). Hebbian Plasticity. In The Handbook of Brain Theory andNeural Networks, M.A. Arbib, ed. (MIT Press), pp. 515–522.
Fregnac, Y. (2012). Reading out the synaptic echoes of low level perception inV1. Lect. Notes Comput. Sci. 7583, 486–495.
Friedrich, R.W., and Laurent, G. (2001). Dynamic optimization of odor repre-sentations by slow temporal patterning of mitral cell activity. Science 291,889–894.
Gilbert, C.D., and Li, W. (2012). Adult visual cortical plasticity. Neuron 75,250–264.
Gregory, R.L. (1997). Knowledge in perception and illusion. Philos. Trans. R.Soc. Lond. B Biol. Sci. 352, 1121–1127.
Grossberg, S., Yazdanbakhsh, A., Cao, Y., and Swaminathan, G. (2008). Howdoes binocular rivalry emerge from cortical mechanisms of 3-D vision? VisionRes. 48, 2232–2250.
Haddad, R., Lanjuin, A., Madisen, L., Zeng, H., Murthy, V.N., and Uchida, N.(2013). Olfactory cortical neurons read out a relative time code in the olfactorybulb. Nat. Neurosci. 16, 949–957.
Haider, B., Krause, M.R., Duque, A., Yu, Y., Touryan, J., Mazer, J.A., and Mc-Cormick, D.A. (2010). Synaptic and network mechanisms of sparse and reli-able visual cortical activity during nonclassical receptive field stimulation.Neuron 65, 107–121.
Hartline, H.K. (1938). The response of single optic nerve fibers of the vertebrateeye to illumination of the retina. Am. J. Physiol. 121, 400–415.
Hein, A., and Held, R. (1967). Dissociation of the visual placing response intoelicited and guided components. Science 158, 390–392.
Held, R., Ostrovsky, Y., de Gelder, B., Gandhi, T., Ganesh, S., Mathur, U., andSinha, P. (2011). The newly sighted fail to match seen with felt. Nat. Neurosci.14, 551–553.
Horton, J.C., and Adams, D.L. (2005). The cortical column: a structure withouta function. Philos. Trans. R. Soc. Lond. B Biol. Sci. 360, 837–862.
Hosoya, T., Baccus, S.A., and Meister, M. (2005). Dynamic predictive codingby the retina. Nature 436, 71–77.
Huxter, J., Burgess, N., and O’Keefe, J. (2003). Independent rate and temporalcoding in hippocampal pyramidal cells. Nature 425, 828–832.
Neuron
Perspective
Isaacson, J.S., and Scanziani, M. (2011). How inhibition shapes cortical activ-ity. Neuron 72, 231–243.
Jack, C.E., and Thurlow, W.R. (1973). Effects of degree of visual associationand angle of displacement on the ‘‘ventriloquism’’ effect. Percept. Mot. Skills37, 967–979.
Jancke, D., Chavane, F., Naaman, S., and Grinvald, A. (2004). Imaging corticalcorrelates of illusion in early visual cortex. Nature 428, 423–426.
Kavukcuoglu, K., Ranzato, M.A., Fergus, R., and LeCun, Y. (2009). Learninginvariant features through topographic filter maps. IEEE Conference on Com-puter Vision and Pattern Recognition, 1605–1612.
Kayser, C., Logothetis, N.K., and Panzeri, S. (2010). Visual enhancement of theinformation representation in auditory cortex. Curr. Biol. 20, 19–24.
Koenderink, J.J. (2012). Geometry of imaginary spaces. J. Physiol. Paris 106,173–182.
Laramee, M.E., and Boire, D. (2014). Visual cortical areas of the mouse: com-parison of parcellation and network structure with primates. Front. Neural Cir-cuits 8, 149.
Laurent, G. (2002). Olfactory network dynamics and the coding of multidimen-sional signals. Nat. Rev. Neurosci. 3, 884–895.
LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learning. Nature 521,436–444.
Lee, T.S., and Nguyen, M. (2001). Dynamics of subjective contour formation inthe early visual cortex. Proc. Natl. Acad. Sci. USA 98, 1907–1911.
Lemus, L., Hernandez, A., Luna, R., Zainos, A., and Romo, R. (2010). Do sen-sory cortices processmore than one sensory modality during perceptual judg-ments? Neuron 67, 335–348.
Leopold, D.A., and Logothetis, N.K. (1996). Activity changes in early visual cor-tex reflect monkeys’ percepts during binocular rivalry. Nature 379, 549–553.
Letzkus, J.J., Kampa, B.M., and Stuart, G.J. (2006). Learning rules for spiketiming-dependent plasticity depend on dendritic synapse location.J. Neurosci. 26, 10420–10429.
Li, W., Piech, V., and Gilbert, C.D. (2006). Contour saliency in primary visualcortex. Neuron 50, 951–962.
Lien, A.D., and Scanziani, M. (2013). Tuned thalamic excitation is amplified byvisual cortical circuits. Nat. Neurosci. 16, 1315–1323.
Liu, B.H., Li, P., Sun, Y.J., Li, Y.T., Zhang, L.I., and Tao, H.W. (2010). Inter-vening inhibition underlies simple-cell receptive field structure in visual cortex.Nat. Neurosci. 13, 89–96.
London, M., Roth, A., Beeren, L., Hausser, M., and Latham, P.E. (2010). Sensi-tivity to perturbations in vivo implies high noise and suggests rate coding incortex. Nature 466, 123–127.
Luczak, A., Bartho, P., and Harris, K.D. (2009). Spontaneous events outline therealm of possible sensory responses in neocortical populations. Neuron 62,413–425.
Machens, C.K.,Wehr, M.S., and Zador, A.M. (2004). Linearity of cortical recep-tive fields measured with natural sounds. J. Neurosci. 24, 1089–1100.
Mante, V., Bonin, V., and Carandini, M. (2008). Functional mechanismsshaping lateral geniculate responses to artificial and natural stimuli. Neuron58, 625–638.
Markov, N.T., Ercsey-Ravasz, M., Van Essen, D.C., Knoblauch, K., Toroczkai,Z., and Kennedy, H. (2013). Cortical high-density counterstream architectures.Science 342, 1238406.
Markram, H. (2012). The human brain project. Sci. Am. 306, 50–55.
Markram, H., and Perin, R. (2011). Innate neural assemblies for lego memory.Front. Neural Circuits 5, 6.
Markram, H., Lubke, J., Frotscher, M., and Sakmann, B. (1997). Regulation ofsynaptic efficacy by coincidence of postsynaptic APs and EPSPs. Science275, 213–215.
Marr, D. (1982). Vision: A Computational Investigation into the Human Repre-sentation and Processing of Visual Information (Freeman).
Martin, K.A., Roth, S., and Rusch, E.S. (2014). Superficial layer pyramidal cellscommunicate heterogeneously between multiple functional domains of catprimary visual cortex. Nat. Commun. 5, 5252.
McFarland, J.M., Cui, Y., and Butts, D.A. (2013). Inferring nonlinear neuronalcomputation based on physiologically plausible inputs. PLoS Comput. Biol.9, e1003143.
McGurk, H., and MacDonald, J. (1976). Hearing lips and seeing voices. Nature264, 746–748.
Mendola, J.D., Dale, A.M., Fischl, B., Liu, A.K., and Tootell, R.B. (1999). Therepresentation of illusory and real contours in human cortical visual areas re-vealed by functional magnetic resonance imaging. J. Neurosci. 19, 8560–8572.
Miikkulainen, R., Bednar, J., Choe, Y., and Sirosh, J. (2006). ComputationalMaps in the Visual Cortex (Springer).
Miller, S.M., Liu, G.B., Ngo, T.T., Hooper, G., Riek, S., Carson, R.G., and Petti-grew, J.D. (2000). Interhemispheric switching mediates perceptual rivalry.Curr. Biol. 10, 383–392.
Miura, K., Mainen, Z.F., and Uchida, N. (2012). Odor representations in olfac-tory cortex: distributed rate coding and decorrelated population activity.Neuron 74, 1087–1098.
Monier, C., Chavane, F., Baudot, P., Graham, L.J., and Fregnac, Y. (2003).Orientation and direction selectivity of synaptic inputs in visual cortical neu-rons: a diversity of combinations produces spike tuning. Neuron 37, 663–680.
Monier, C., Fournier, J., and Fregnac, Y. (2008). In vitro and in vivomeasures ofevoked excitatory and inhibitory conductance dynamics in sensory cortices.J. Neurosci. Methods 169, 323–365.
Montaser-Kouhsari, L., Landy, M.S., Heeger, D.J., and Larsson, J. (2007).Orientation-selective adaptation to illusory contours in human visual cortex.J. Neurosci. 27, 2186–2195.
Morel, J.M., Petro, A.B., and Sbert, C. (2010). A PDE formalization of Retinextheory. IEEE Trans. Image Process. 19, 2825–2837.
Moser, E.I., and Moser, M.B. (2013). Grid cells and neural coding in high-endcortices. Neuron 80, 765–774.
Movshon, J.A. (2013). Functional models and functional organisms for sys-tems neuroscience. In Cosyne (Salt Lake City), p. 25.
Mumford, D. (1991). On the computational architecture of the neocortex. I. Therole of the thalamo-cortical loop. Biol. Cybern. 65, 135–145.
Musall, S., von der Behrens, W., Mayrhofer, J.M., Weber, B., Helmchen, F.,and Haiss, F. (2014). Tactile frequency discrimination is enhanced by circum-venting neocortical adaptation. Nat. Neurosci. 17, 1567–1573.
Nauhaus, I., Benucci, A., Carandini, M., and Ringach, D.L. (2008). Neuronalselectivity and local map structure in visual cortex. Neuron 57, 673–679.
Niell, C.M., and Stryker, M.P. (2008). Highly selective receptive fields in mousevisual cortex. J. Neurosci. 28, 7520–7536.
O’Connor, D.H., Hires, S.A., Guo, Z.V., Li, N., Yu, J., Sun, Q.Q., Huber, D., andSvoboda, K. (2013). Neural coding during active somatosensation revealed us-ing illusory touch. Nat. Neurosci. 16, 958–965.
O’Keefe, J., and Recce, M.L. (1993). Phase relationship between hippocampalplace units and the EEG theta rhythm. Hippocampus 3, 317–330.
Ohki, K., Chung, S., Kara, P., Hubener, M., Bonhoeffer, T., and Reid, R.C.(2006). Highly ordered arrangement of single neurons in orientation pinwheels.Nature 442, 925–928.
Olcese, U., Iurilli, G., and Medini, P. (2013). Cellular and synaptic architectureof multisensory integration in the mouse neocortex. Neuron 79, 579–593.
Palacios, A.G., and Bozinovic, F. (2003). An ‘‘enactive’’ approach to integrativeand comparative biology: thoughts on the table. Biol. Res. 36, 101–105.
Neuron 88, October 7, 2015 ª2015 Elsevier Inc. 125
Neuron
Perspective
Pan, Y., Chen, M., Yin, J., An, X., Zhang, X., Lu, Y., Gong, H., Li, W., andWang,W. (2012). Equivalent representation of real and illusory contours in macaqueV4. J. Neurosci. 32, 6760–6770.
Pillow, J.W., Shlens, J., Paninski, L., Sher, A., Litke, A.M., Chichilnisky, E.J.,and Simoncelli, E.P. (2008). Spatio-temporal correlations and visual signallingin a complete neuronal population. Nature 454, 995–999.
Polat, U., and Sagi, D. (1993). Lateral interactions between spatial channels:suppression and facilitation revealed by lateral masking experiments. VisionRes. 33, 993–999.
Polat, U., Mizobe, K., Pettet, M.W., Kasamatsu, T., and Norcia, A.M. (1998).Collinear stimuli regulate visual responses depending on cell’s contrastthreshold. Nature 391, 580–584.
Priebe, N.J., and Ferster, D. (2012). Mechanisms of neuronal computation inmammalian visual cortex. Neuron 75, 194–208.
Rabinovich, M., Huerta, R., and Laurent, G. (2008). Neuroscience. Transientdynamics for neural processing. Science 321, 48–50.
Raposo, D., Kaufman, M.T., and Churchland, A.K. (2014). A category-free neu-ral population supports evolving demands during decision-making. Nat. Neu-rosci. 17, 1784–1792.
Ratliff, F. (1965). Mach Bands: Quantitative Studies on Neural Networks in theRetina (Holden-Day).
Reinagel, P., and Reid, R.C. (2000). Temporal coding of visual information inthe thalamus. J. Neurosci. 20, 5392–5400.
Riehle, A., Grun, S., Diesmann, M., and Aertsen, A. (1997). Spike synchroniza-tion and rate modulation differentially involved in motor cortical function. Sci-ence 278, 1950–1953.
Rust, N.C., Schwartz, O., Movshon, J.A., and Simoncelli, E.P. (2005). Spatio-temporal elements of macaque v1 receptive fields. Neuron 46, 945–956.
Rust, N.C., Mante, V., Simoncelli, E.P., and Movshon, J.A. (2006). How MTcells analyze the motion of visual patterns. Nat. Neurosci. 9, 1421–1431.
Safo, P.K., and Regehr, W.G. (2005). Endocannabinoids control the inductionof cerebellar LTD. Neuron 48, 647–659.
Sceniak, M.P., Ringach, D.L., Hawken, M.J., and Shapley, R. (1999). Con-trast’s effect on spatial summation by macaque V1 neurons. Nat. Neurosci.2, 733–739.
Seghier, M.L., and Vuilleumier, P. (2006). Functional neuroimaging findings onthe human perception of illusory contours. Neurosci. Biobehav. Rev. 30,595–612.
Sharma, J., Angelucci, A., and Sur, M. (2000). Induction of visual orientationmodules in auditory cortex. Nature 404, 841–847.
Sharma, J., Dragoi, V., Tenenbaum, J.B., Miller, E.K., and Sur, M. (2003). V1neurons signal acquisition of an internal representation of stimulus location.Science 300, 1758–1763.
Sheth, B.R., Sharma, J., Rao, S.C., and Sur, M. (1996). Orientation maps ofsubjective contours in visual cortex. Science 274, 2110–2115.
Shulz, D.E., and Fregnac, Y. (2010). From sensation to perception. In Encyclo-pedia of Behavioural Neuroscience, G. Koob, M. Le Moal, and R. Thompson,eds. (Elsevier), pp. 550–558.
Sillito, A.M., Grieve, K.L., Jones, H.E., Cudeiro, J., and Davis, J. (1995). Visualcortical mechanisms detecting focal orientation discontinuities. Nature 378,492–496.
Smyth, D.,Willmore, B., Baker, G.E., Thompson, I.D., and Tolhurst, D.J. (2003).The receptive-field organization of simple cells in primary visual cortex of fer-rets under natural scene stimulation. J. Neurosci. 23, 4746–4759.
Solomon, S.G., Lee, B.B., and Sun, H. (2006). Suppressive surrounds andcontrast gain in magnocellular-pathway retinal ganglion cells of macaque.J. Neurosci. 26, 8715–8726.
Sompolinsky, H., and Shapley, R. (1997). New perspectives on the mecha-nisms for orientation selectivity. Curr. Opin. Neurobiol. 7, 514–522.
126 Neuron 88, October 7, 2015 ª2015 Elsevier Inc.
Spinelli, D.N., and Jensen, F.E. (1979). Plasticity: the mirror of experience. Sci-ence 203, 75–78.
Spors, H., Wachowiak, M., Cohen, L.B., and Friedrich, R.W. (2006). Temporaldynamics and latency patterns of receptor neuron input to the olfactory bulb.J. Neurosci. 26, 1247–1259.
Stein, B.E., and Stanford, T.R. (2008). Multisensory integration: current issuesfrom the perspective of the single neuron. Nat. Rev. Neurosci. 9, 255–266.
Stopfer, M., Jayaraman, V., and Laurent, G. (2003). Intensity versus identitycoding in an olfactory system. Neuron 39, 991–1004.
Super, H., van der Togt, C., Spekreijse, H., and Lamme, V.A. (2003). Internalstate of monkey primary visual cortex (V1) predicts figure-ground perception.J. Neurosci. 23, 3407–3414.
Szabo, V., Ventalon, C., De Sars, V., Bradley, J., and Emiliani, V. (2014).Spatially selective holographic photoactivation and functional fluorescenceimaging in freely behaving mice with a fiberscope. Neuron 84, 1157–1169.
Theunissen, F.E., and Elie, J.E. (2014). Neural processing of natural sounds.Nat. Rev. Neurosci. 15, 355–366.
Troyer, T.W., Krukowski, A.E., Priebe, N.J., and Miller, K.D. (1998). Contrast-invariant orientation tuning in cat visual cortex: thalamocortical input tuningand correlation-based intracortical connectivity. J. Neurosci. 18, 5908–5927.
Vaadia, E., Haalman, I., Abeles, M., Bergman, H., Prut, Y., Slovin, H., and Aer-tsen, A. (1995). Dynamics of neuronal interactions in monkey cortex in relationto behavioural events. Nature 373, 515–518.
Van Rullen, R., and Thorpe, S.J. (2001). Rate coding versus temporal ordercoding: what the retinal ganglion cells tell the visual cortex. Neural Comput.13, 1255–1283.
Vidyasagar, T.R., and Eysel, U.T. (2015). Origins of feature selectivities andmaps in the mammalian primary visual cortex. Trends Neurosci. 38, 475–485.
Vinje, W.E., and Gallant, J.L. (2000). Sparse coding and decorrelation in pri-mary visual cortex during natural vision. Science 287, 1273–1276.
von der Heydt, R., Peterhans, E., andBaumgartner, G. (1984). Illusory contoursand cortical neuron responses. Science 224, 1260–1262.
Von der Malsburg, C. (1981). The correlation theory of brain function (Max-Plank Institute for Biophysical Chemistry).
von der Malsburg, C. (1986). Am I thinking brain assemblies? In Brain Theory,G. Palm and A. Aertsen, eds. (Springer), pp. 161–176.
Wagemans, J., Elder, J.H., Kubovy, M., Palmer, S.E., Peterson, M.A., Singh,M., and von der Heydt, R. (2012). A century of Gestalt psychology in visualperception: I. Perceptual grouping and figure-ground organization. Psychol.Bull. 138, 1172–1217.
Wallace, D.J., Greenberg, D.S., Sawinski, J., Rulla, S., Notaro, G., and Kerr,J.N. (2013). Rats maintain an overhead binocular field at the expense of con-stant fusion. Nature 498, 65–69.
Watkins, S., Shams, L., Josephs, O., and Rees, G. (2007). Activity in human V1follows multisensory perception. Neuroimage 37, 572–578.
Wehr, M., and Zador, A.M. (2005). Synaptic mechanisms of forward suppres-sion in rat auditory cortex. Neuron 47, 437–445.
Wertheimer, M. (1912). Experimentelle Studien uber das Sehen von Bewe-gung. In Zeitschrift fur Psychologie (J.A. Barth).
Yeh, C.I., Xing, D., Williams, P.E., and Shapley, R.M. (2009). Stimulusensemble and cortical layer determine V1 spatial receptive fields. Proc. Natl.Acad. Sci. USA 106, 14652–14657.
Zingg, B., Hintiryan, H., Gou, L., Song, M.Y., Bay, M., Bienkowski, M.S., Fos-ter, N.N., Yamashita, S., Bowman, I., Toga, A.W., and Dong, H.W. (2014). Neu-ral networks of the mouse neocortex. Cell 156, 1096–1111.
Zipser, K., Lamme, V.A., and Schiller, P.H. (1996). Contextual modulation inprimary visual cortex. J. Neurosci. 16, 7376–7389.