Lecture5

Understanding Biological Autonomy

Takashi Ikegami University of Tokyo

1

What is Biological autonomy?

“Fluctuation of sensorimotor performance”, “Capricious behaviors”, “Mind-Wandering”, “Stimulus Independent Thoughts”,,,,

These are the phrases that characterize “biological autonomy”, which we think distinguish life from man-made machines. But the biological autonomy can’t be simply attributed to the internal dynamics of a system. We need to many different ways of realizing embodiment; chemical, physical and computational embodiment to understand the biological autonomy.

2

Computational embodimentby Keisuke Suzuki and T.Ikegami (2005)

Physical embodiment Grey Walter’s turtle robots (1951)

Physical embodiment Music robots “Miuro” by ZMP (2007)

Chemical embodiment Hanczyc et al. (2007)

3

+

“hard” part:sensori-motorcouplinge.g. chemotaxis

Noise part(unwanted part)

A general picture (robot hypothesis):

sensory pattern

motor output

= an autonomous agent

4

Our picture (“Variation Generator” hypothesis):

“hard” sensorimotor coupling

fluctuating background dynamics as a generator of variations, which leads to embodied chaotic itinerancy

=　An individual

sensory pattern

motor output

crystallizing “rule”

5

Quantifying, Designing and Synthesizing biological autonomy

i) Quantifying housefly’s autonomous behavior (real life) Takahashi, H. Horibe, N, Ikegami, T and Shimada, M, ”Analyzing House Fly’s Exploration Behaviorwith AR methods” (submitted to JPSJ, physics/0702170).

ii) Theory of Autonomy (computational embodiment) Ikegami, T. Simulating Active Perception and Mental Imagery with Embodied Chaotic Itinerancy, J.Consciousness Studies Vol.14 (2007) ppp.111-125.)

iii)Designing Autonomy with zmp robot (physical embodiment) Making a Robot Dance to Music Using Chaotic Itinerancy in a Network of FitzHugh-Nagumo Neurons” by Jean-Julien Aucouturier, Yuta Ogai, Takashi Ikegami

iv)Emergence of autonomy in oil droplet (chemical embodiment) Hanczyc, M., Toyota,T., Ikegami, T., Packard, N. and Sugawara, T. " Chemistry at the oil-water interface: Self-propelled oil droplets"J. Am. Chem. Soc.; (2007); 129(30) pp 9386 - 939.

6

7

Autonomy in Exploring Behavior of flies

Takahashi, H. Horibe, N, Ikegami, T and Shimada, M, ”Analyzing a House Fly’s Exploration Behavior with AutoRegression methods” (submitted to JPSJ, physics/0702170).

7

9

Temporal evolution of AR dimensions

velocity

heading angle

9

10

(a) with sugar droplets (b) without sugar droplets

Anomalous diffusion? w/o sugar droplets

< x > = ta=1.55

2 a

10

11

Random walks and Levy flights observed in fluid flowsER Weeks and HL Swinney, formerly online at Nonlinear Science Today (Springer-Verlag, 1998).

11

Coupling/Decoupling phenomena:

Swinderen, V. Bruno, Nitz, A. Douglas, Greenspan, J. Ralph (2004), ‘Uncouplingof brain activity from movement defines arousal states in Drosophila’, CurrentBiology, 14, pp.81–87

The correlation between local field potentials (LFPs) in the brain and overt movements of the fruit fly during different states of arousal, such as spontaneous daytime waking movement, visual arousal, spontaneous night-time movement, and stimulus-induced movement.

12

These experiments indicate that the relationship between brain LFPs and movement in the fruit fly is dynamic and that the degree of coupling between these two measures of activity defines distinct states of arousal. It has been observed that neural activities (20-30Hz) can be coupled/decoupled from the apparent bodily movement.

13

Output Neuron (2+2)

Input Neuron (10)

Internal Neuron (16)

There are two kinds of neurons: One that emits fast pulse signals and one emits slow pulse signals.　

OBJECT

Theory of Autonomy: Ikegami, T. Simulating Active Perception and Mental Imagery with Embodied Chaotic Itinerancy, J.Consciousness Studies Vol.14 (2007) ppp.111-125.

14

time (t)

Integrated activity

input neuron activities

internal neuron activities

Time evolution of neural states

15

e) 900-1000 time steps f) 1100-1200 time steps

g) 1500-1600 time steps h) 1600-1700 time steps

Fig.3. Time Evolution of navigation patterns. A structure of motion pattern switches from one to the other. From the left upper corner (a) to the right bottom corner (h), spatial trails are overlaid for each 100 time steps.

It sometimes happens that the navigation patterns can't activate input neu-rons. Yet, the navigation style changes from time to time. This is essentially caused by the chaotic itinerant dynamics within the internal network.Chaotic and periodic local attractors can be distinguished by drawing the re-turn map of the internal neural dynamics. Here I draw a 2-dimensional return map of the internal and output neurons. A clear limit cycle is obtained for the local periodic attractor and a noisy boundary is observed in the chaotic one. A chaotic itinerant case of no external input is given with its navigation pattern. In this CI case, the return map shows a clear evidence of a stretch-ing and folding processes. By changing the spatial scales of checker-board pattern, we notice that only a certain spatial pattern causes navigation mo-tion. This is well understood since the spatial scale determines the temporal interval of successive stimulus and the navigation style is highly sensitive to

Fig.2. Itinerant behavior of the integrated input and internal neural activities. Different temporal pattern corresponds to the different local attractor, which is associated with the spatial patterns in Fig.3. It should be noted that the spiky patterns of the input neurons are corresponding to the chaotic internal activities. The set of parameters are, w=0.1, I=0.7, a=0.7, b=0.8 and C=10. When an internal neural state falls down to a periodic attractor, its motor out-put becomes periodic so that the navigation trail also becomes periodic. If no new inputs are given from the environment, the agent's periodic motion is eventually preserved. But if the state is chaotic, the periodic motion often becomes unstable and new motion structure appears without having external inputs. A general situation is in between the two cases. Input neurons don't fire against excess pulse trains. They only fire when the pulse trains have some accents. Namely, the input trains　 should have a contrast between quiet and busy phase. Continual spiking time series can't let neurons fire. Those input patterns are indirectly controlled by the agent's navigation pat-tern. Therefore, some navigation patterns generate input signals that can cause the input neurons to fire, which leads to changed navigation patterns. In this way, chaotic itinerant behavior is generated by the sensory-motor cou-pling.

a) 100-200 time steps b) 500-600 time steps

c) 600-700 time steps d) 800-900 time steps

e) 900-1000 time steps f) 1100-1200 time steps

g) 1500-1600 time steps h) 1600-1700 time steps

Fig.3. Time Evolution of navigation patterns. A structure of motion pattern switches from one to the other. From the left upper corner (a) to the right bottom corner (h), spatial trails are overlaid for each 100 time steps.

It sometimes happens that the navigation patterns can't activate input neu-rons. Yet, the navigation style changes from time to time. This is essentially caused by the chaotic itinerant dynamics within the internal network.Chaotic and periodic local attractors can be distinguished by drawing the re-turn map of the internal neural dynamics. Here I draw a 2-dimensional return map of the internal and output neurons. A clear limit cycle is obtained for the local periodic attractor and a noisy boundary is observed in the chaotic one. A chaotic itinerant case of no external input is given with its navigation pattern. In this CI case, the return map shows a clear evidence of a stretch-ing and folding processes. By changing the spatial scales of checker-board pattern, we notice that only a certain spatial pattern causes navigation mo-tion. This is well understood since the spatial scale determines the temporal interval of successive stimulus and the navigation style is highly sensitive to

Associated navigation pattern

16

A message from this theoretical work:

Embodied Chaotic Itinerancy is an inevitable outcome of coupled unstable neural dynamics. The outcome of such unstable dynamics is the attachment-detachment switching described here.When an agent is attached to the environment, it intends to process the sensory data flow. When it is detached, it obeys its internal dynamics. We claim that conscious states emerge when a subject spontaneously selects one of the two phases.

17

Making a Robot Dance to Music Using Chaotic Itinerancy in a Network of FitzHugh-Nagumo Neurons” by Jean-Julien Aucouturier, Yuta Ogai, Takashi Ikegami (ICONIP2007)

18

Luisi’s special soup(Micelles + water )

+Oil droplet( oleate anhydride (precursor of surfactant)+ nitrobenzene)

+High pH ( >11)=?

Oleic acid

Oleic anhydride

At room temp

2

H2O

Chemical EmbodimentChemistry at the oil-water interface: Self-propelled oil droplets

Hanczyc,M. Toyota,T, Ikegami,T. Packard,N. and Sugawara, T. J. Am. Chem. Soc.; (2007); 129(30) pp 9386 - 939.

20

Here is an example of self-running oil droplet. (thanks to Martin Hanczyc)

The system consists simply of an oil droplet in an aqueous environment. The aqueous phase contains a surfactant that modulates the interfacial tension between the drop of oil and its environment. We embed a chemical reaction in the oil phase that reacts with water when an oily precursor comes in contact with the water phase at the liquid-liquid interface.

21

this movie is due to Martin Hanczyc.(ProtLife s.r.l.)the size of the droplet is a few 100 um.

This reaction not only powers the droplet to move in the aqueous phase but also allows for sustained movement. The direction of the movement is governed by a self-generated pH gradient that surrounds the droplet. In addition this self-generated gradient can be overridden by an externally imposed pH gradient, and therefore the direction of droplet motion may be controlled.

22

An internal convection flow is observed.

Fluorescent micrograph of a self-moving oil droplet with internal convection. Visible are the dye-filled internal compartments that formed and moved convectively within the oil droplet. The size of this droplet is nearly 0.3mm in diameter.

Also we noticed that convection flow is generated inside the oil droplet to cause the movement. We can observe that the droplet senses the gradient in the environment (either internally generated or externally imposed) and moves predictably within the gradient as a form of primitive chemotaxis.

A possible mechanism of the self-movement was studied with the Navier-Stokes equation with the chemical reaciton. see Hiroki Matsuno, Martin M. Hanczyc, and Takashi Ikegami, “ Self-maintained Movements of Droplets with Convection Flow” in the proceedings of the 3rd Australian Artificial Conference (2007).

23

24

24

24

Summary:

Real Life: Quantification of autonomy is given by the spontaneous alternation of AR dimension and the anomalous diffusion with a fly’s exploring behavior.

Computational embodiment: We then showed that embodied chaotic itinerancy (ECI) can be used for designing an autonomous robot (physical embodiment).

Finally we showed that the simple autonomy emerges from chemical embodiment.

24

24

Summary:




Yes. But see “Self-Motile Colloidal Particles: From Directed Propulsion to RandomWalk” Jonathan R. Howse,1 Richard A. L. Jones,* Anthony J. Ryan,Tim Gough, Reza Vafabakhsh, and Ramin Golestanian

24

24

Summary:





But the idea of computational embodiment needs to be elaborated further.

24

24

Summary:





But the idea of computational embodiment needs to be elaborated further.

We need more behavior variations for biological autonomy. Also recycling of energy is a big problem here.

24

Lecture5

Education

Transcript of Lecture5