Intelligence Without Representation

1
Intelligence Without Representation

• Review of Rodney Brooks article
• Presented by Teo Susnjak

2
Background

• Wtitten in 1987
• The backdrop to this paper is the struggle for AI
  research to get past the issue of representation
• AI started off with goal to replicate human level
  intelligence in a machine
• Brooks claim is that we need to first learn how
  to practice with lower level intelligence before
  attempting to decompose human level intelligence
  into sub pieces and figure out the interfaces
  between them.
• Whats Brooks approach to developing intelligent
  systems?
• Incremental build up of capabilities having a
  complete system at each step
• at each step the complete intelligent system is
  let loose in the real world
• What are the conclusions reached?
• Explicit representations and models of the world
  get in the way when examining simple intelligence
• Whats the hypothesis?
• Representation is the wrong unit of abstraction
  in building the bulkiest part of intelligent
  systems

3
Evolution of intelligence

• If the theory of evolution correctly explains our
  existence then its assumed
• 3.5 billion years ago first single cells
• 1 billion years later photosynthetic plants
• 1.5 billion years later first fish arrived
• 100 million years later insect
• 80 million years later reptiles
• 40 million years later dinosaurs
• 130 million years later primates
• 102 million years later great apes
• 15.5 million years later, 2.5 million years ago
  humans
• agriculture invented 100,000 years ago
• 5000 years ago writing
• expert knowledge in the last few hundred years
• What conclusions can we draw from this about the
  difficulty of intelligence?
• Problem solving and expert knowledge come fairly
  easily once mobility and functioning in a dynamic
  environment figured out.
• Mobility, vision, and ability to carry out
  survival tasks in a changing environment provide
  a basis for the development of true intelligence.

4
Abstraction

• Why is it hard for AI to succeed?
• Whenever AI researcher solve a supposed AI
  problem with an algorithm it is later claimed
  that it was never an AI problem in the first
  place
• Why is it easy for AI to succeed?
• Typically AI defines problems that are not solved
  as non-AI problems.
• What principal mechanism for this partitioning of
  problems is used?
• Abstraction factoring out all aspects of
  perception and motor skills.
• How did the idea that representation is the key
  problem in AI arise?
• The block world in the 60s and 70s
• key to success was to represent the world
  completely and explicitly
• it was a toy world in the end and the daunting
  task of representing everything in a real world
  became apparent

5
Abstraction

• What is abstraction?
• Abstraction is the process of generalization by
  reducing the information content of a concept or
  an observable phenomenon, typically in order to
  retain only information which is relevant for a
  particular purpose
• Why is this a real problem in AI?
• At present the abstraction is carried out by the
  programmers and all AI programs have to do is
  search.
• The challenge is to have the program itself
  perform the abstraction.

6
Incremental Intelligence

• Brooks notion of Creatures
• completely autonomous mobile agents that co-exist
  in the world with humans and are seen by them as
  intelligent beings themselves.
• What are the requirements?
• cope and act timely in a changing environment
• robust with changes in environment. Minor
  degradation expected as changes increase in the
  environment.
• Maintain multi-goals and change them as required
  making it adaptable
• it should have a purpose

7
Incremental Intelligence

• What are the features of decomposition by
  function?
• Central system, perceptual modules and action
  modules
• Strengths?
• We can brake up bulky parts like the central
  system into knowledge representation, learning,
  planning, qualitative reasoning etc.
• Weaknesses?
• Bugs hard to fix
• long chain of modules between perception and
  action and to test them, they all must be built.
• What are the features of decomposition by
  activity?
• Each activity connects sensing to action directly
• Strengths?
• A clear incremental path for simple to complex
  systems. Easy to add behaviours.
• Weaknesses?
• There may be a limit as to how many layers can
  be added

8
Representations

• Where are the representations in this system?
• Output from perception cannot be located anywhere
• there is simultaneous processing of sensing data
  at different levels
• What is the common theme?
• There is no central representation system
• Arent some of the internal variables
  representations?
• There are no variables that need instantiation in
  the reasoning process
• the state of the world determines the action of
  the creature

9
Subsumption Methodology

• Description of the methodology
• The Creatures must be tested with each
  additional layer in the real world.
• Why not just test in a test environment?
• Easy to build a system which relies on simplified
  representations of the world. This reliance can
  then be propagated to other modules and then the
  entire system needs to be rebuilt.
• When adding layers where could the bugs be found?
• In the current layer being added.
• The interactions between layers
• the only layer that can be modified to fix the
  bugs even if theyre in lower layers is the
  current layer
• Application
• 4 robots were built and operate in a real dynamic
  world
• all pursue multiple goals
• all have multi layers and a FSM on each one
  running asynchronously
• no central control
• 3 layers
• 1. Avoid objects
• 2. Wander
• 3. Explore distant places

10
Subsumption Methodology

• Is it Connectionism or Neural networks?
• In subsumption unlike connectionism the nodes are
  unique FSM and their connections are not dense,
  non-uniform and low between layers.
• Neural networks claim some parallels to
  biological systems whereas subsumption doesnt
• Is it Production rules or knowledge based?
• No since there are no rule matchings since there
  is no rule base.
• Not knowledge based since everything is
  hard-wired
• How large can it become?
• Seems very expandable
• Is learning possible?
• A fixed topology networks of the FSM which can
  perform learning have been built

11
Artificial Flight Analogy

• Artificial flight analogy.
• What insight does it give into the field of AI?
• The challenge of trying to create a complete AI
  system all at once especially when the individual
  sub parts and their interactions are not well
  understood.

12
Conclusions

• Building complex AI systems all at once is very
  difficult
• an incremental approach appears more favourable
• the incremental systems should function in the
  real world
• representations and models of the real world get
  in the way
• the AI systems themselves should be able to
  perform abstraction

13

• Connectionism is an approach in the fields of
  artificial intelligence, cognitive
  psychology/cognitive science, neuroscience and
  philosophy of mind. Connectionism models mental
  or behavioral phenomena as the emergent processes
  of interconnected networks of simple units. There
  are many different forms of connectionism, but
  the most common forms utilize neural network
  models
• The central connectionist principle is that
  mental phenomena can be described by
  interconnected networks of simple units. The form
  of the connections and the units can vary from
  model to model. For example, units in the network
  could represent neurons and the connections could
  represent synapses. Another model might make each
  unit in the network a word, and each connection
  an indication of semantic similarity.
• NeuralTraditionally, the term neural network had
  been used to refer to a network of biological
  neurons. In more common usage, the term is often
  used to refer to artificial neural networks,
  which are composed of artificial neurons or
  nodes. Thus the term 'Neural Network' has two
  distinct connotationsBiological neural networks
  are made up of real biological neurons that are
  connected or functionally-related in the
  peripheral nervous system or the central nervous
  system. In the field of neuroscience, they are
  often identified as groups of neurons that
  perform a specific physiological function in
  laboratory analysis.Artificial neural networks
  are made up of interconnecting artificial neurons
  (usually simplified neurons) which may share some
  properties of biological neural networks.
  Artificial neural networks may either be used to
  gain an understanding of biological neural
  networks, or for solving traditional artificial
  intelligence tasks without necessarily attempting
  to model a real biological system.Please see the
  corresponding articles for details on artificial
  neural networks or biological neural networks.
  This article focuses on the relationship between
  the two concepts.editCharacterizationIn general
  a biological neural network is composed of a
  group or groups of physically connected or
  functionally associated neurons. A single neuron
  can be connected to many other neurons and the
  total number of neurons and connections in a
  network can be extremely large. Connections,
  called synapses, are usually formed from axons to
  dendrites, though dendrodendritic microcircuits
  1 and other connections are possible. Apart
  from the electrical signaling, there are other
  forms of signaling that arise from
  neurotransmitter diffusion, which have an effect
  on electrical signaling. As such, neural networks
  are extremely complex. While a detailed
  description of neural systems seems currently
  unattainable, progress is being made towards a
  better understanding of basic mechanisms.