Intelligence Without Representation€¦ · Representation Author: Rodney Brooks Presenters: Alex...

18
Intelligence Without Representation Author: Rodney Brooks Presenters: Alex Kuefler & Rachel Luo

Transcript of Intelligence Without Representation€¦ · Representation Author: Rodney Brooks Presenters: Alex...

Page 1: Intelligence Without Representation€¦ · Representation Author: Rodney Brooks Presenters: Alex Kuefler & Rachel Luo. Background But Deep Representation Learning is one of many

Intelligence Without Representation

Author: Rodney BrooksPresenters: Alex Kuefler & Rachel Luo

Page 2: Intelligence Without Representation€¦ · Representation Author: Rodney Brooks Presenters: Alex Kuefler & Rachel Luo. Background But Deep Representation Learning is one of many

BackgroundBut Deep Representation Learning is one of many thrusts in AI.

Neural networks have been going in and out of style since 1950s.

State-of-the-Art throughout decades has differed not only in modelling choices, but in design philosophies.

Page 3: Intelligence Without Representation€¦ · Representation Author: Rodney Brooks Presenters: Alex Kuefler & Rachel Luo. Background But Deep Representation Learning is one of many

BackgroundIntelligence Without Representation (1987) is an abridged manifesto by Rodney Brooks.

Paper presents an orthogonal view of thinking about (artificial) intelligence.

Not state of the art representation learning, but neural networks were recently resurrected as well.

Page 4: Intelligence Without Representation€¦ · Representation Author: Rodney Brooks Presenters: Alex Kuefler & Rachel Luo. Background But Deep Representation Learning is one of many

Outline1. Introduction2. Critique of Pure Representation3. Subsumption Architecture4. The Methodology, In Practice5. Going Forward

Page 5: Intelligence Without Representation€¦ · Representation Author: Rodney Brooks Presenters: Alex Kuefler & Rachel Luo. Background But Deep Representation Learning is one of many

IntroductionArtificial Intelligence was first conceived to recreate human intelligence.

Later work focused instead on subdomains: NLP, Vision, Games, etc.

Division of labor is great for applications, but are we still moving towards general AI?

Brooks says “No”

Page 6: Intelligence Without Representation€¦ · Representation Author: Rodney Brooks Presenters: Alex Kuefler & Rachel Luo. Background But Deep Representation Learning is one of many

Introduction

Shakey: Camera, laser range-finder, grid-world model, A* planning

Roomba: Follows walls. Bounces off walls.

Page 7: Intelligence Without Representation€¦ · Representation Author: Rodney Brooks Presenters: Alex Kuefler & Rachel Luo. Background But Deep Representation Learning is one of many

Critique of Pure RepresentationAI research typically factors problems into two components: Experimenters solve one problem, the AI system solves the other.

Typically, we Abstract Out perception and motor skills from our simulators.

Example: Feature Engineering

Before deep learning, it was the scientist’s job to preprocess raw data (we still do this!)

Page 8: Intelligence Without Representation€¦ · Representation Author: Rodney Brooks Presenters: Alex Kuefler & Rachel Luo. Background But Deep Representation Learning is one of many

Subsumption Architecture

The world is its own representation.

Page 9: Intelligence Without Representation€¦ · Representation Author: Rodney Brooks Presenters: Alex Kuefler & Rachel Luo. Background But Deep Representation Learning is one of many

Subsumption Architecture

Decomposition by Function

Each subsystem handles one transformation from some symbol to another.

Page 10: Intelligence Without Representation€¦ · Representation Author: Rodney Brooks Presenters: Alex Kuefler & Rachel Luo. Background But Deep Representation Learning is one of many

Subsumption Architecture

Decomposition by Activity

Each subsystem produces a complete activity. It decides when to act itself.

Page 11: Intelligence Without Representation€¦ · Representation Author: Rodney Brooks Presenters: Alex Kuefler & Rachel Luo. Background But Deep Representation Learning is one of many

The Methodology, In PracticeMethodological maxims:

1. Test the Creatures in the real world.2. Test each layer as it is built.

To demonstrate this methodology, Brooks et al built four robots based on the principles of task decomposition!

Page 12: Intelligence Without Representation€¦ · Representation Author: Rodney Brooks Presenters: Alex Kuefler & Rachel Luo. Background But Deep Representation Learning is one of many

The Methodology, In PracticeArchitecture:

Subsumption architecture embodies the idea of decomposition by activity into task-achieving layers.

Each layer is comprised of a fixed-topology network of finite state machines.

Layers are combined through mechanisms called suppression and inhibition.

● Suppression: Messages arriving on the new layer suppress messages incoming from existing layers

● Inhibition: Messages on the new layer inhibit messages being emitted on existing layers

Page 13: Intelligence Without Representation€¦ · Representation Author: Rodney Brooks Presenters: Alex Kuefler & Rachel Luo. Background But Deep Representation Learning is one of many

The Methodology, In PracticeExample: Three-layered mobile robot

1. Layer 1: Avoid hitting objects2. Layer 2: Wander around3. Layer 3: Explore new places

Page 14: Intelligence Without Representation€¦ · Representation Author: Rodney Brooks Presenters: Alex Kuefler & Rachel Luo. Background But Deep Representation Learning is one of many

What This is Not

1. Connectionism2. Neural networks3. Production rules4. Blackboard5. German philosophy

Page 15: Intelligence Without Representation€¦ · Representation Author: Rodney Brooks Presenters: Alex Kuefler & Rachel Luo. Background But Deep Representation Learning is one of many

Incremental IntelligenceLike deep learning hierarchical composition obviates need for central representations.

Unlike deep learning the purpose of the system is implicit, and only emerges from the behavior of submodules, whose purposes are explicit.

Page 16: Intelligence Without Representation€¦ · Representation Author: Rodney Brooks Presenters: Alex Kuefler & Rachel Luo. Background But Deep Representation Learning is one of many

Limits to Growth1. How many layers?- So far, three on a physical robot

(six in simulation)2. How complex?- Working on a robot that will require

fourteen layers- Creature that can retrieve empty

soda cans from cluttered desks3. Is learning possible?- Hopefully! Want to achieve

insect-level intelligence

Page 17: Intelligence Without Representation€¦ · Representation Author: Rodney Brooks Presenters: Alex Kuefler & Rachel Luo. Background But Deep Representation Learning is one of many

Further Reading ● Herbert, the soda can-collecting robot:

https://dspace.mit.edu/handle/1721.1/6483● Other advanced robots:

https://people.csail.mit.edu/brooks/papers/fast-cheap.pdf● Cog, the humanoid robot:

http://people.csail.mit.edu/brooks/papers/CMAA-group.pdf● More about Brooks’s manifesto: Flesh and Machines, Pantheon Books, New

York, NY, 2002.

Page 18: Intelligence Without Representation€¦ · Representation Author: Rodney Brooks Presenters: Alex Kuefler & Rachel Luo. Background But Deep Representation Learning is one of many

Bibliography1. Brooks, R. A. (1991). Intelligence without representation. Artificial intelligence, 47(1), 139-159.2. Domingos, P. (2016). Master Algorithm. Penguin Books.3. Dosovitskiy, A., Springenberg, J., Tatarchenko, M., & Brox, T. (2016). Learning to Generate Chairs,

Tables and Cars with Convolutional Networks.4. Nilsson, N. J. (2009). The quest for artificial intelligence. Cambridge University Press.