Associative Learning for a Robot Intelligence


by John H Andreae


Imperial College Press, 1998



  The explanation of brain functioning in terms of the association of ideas has been popular since the 17th century. However, recently the process of association has been dismissed as computationally inadequate by prominent cognitive scientists. In this book, a sharper definition of the term "association" is used to revive the process by showing that associative learning can indeed be computationally powerful. With an appropriate organization, associative learning can be embodied in a robot to realize a human-like intelligence, which sets its own goals, exhibits unique unformalizable behaviour and has no hidden homunculi.



  A series of simulated robot experiments demonstrates Turing Machine power, a simple form of internal "thinking", incipient communication of intentions, control of a hierarchical task, the emergence of cooperation, and some preliminary steps towards the learning of language. The "innate" endowment of the robot includes the association templates, which drive the associative learning, and a new form of short-term visual memory.



  Some say that Artificial Intelligence is undergoing a paradigm shift. Certainly there are several competing ideas and ideals. Neural networks and dynamic systems are offered as alternatives to the information processing and digital computer models of the brain. We are asked to decide between symbolic and subsymbolic, between algorithmic and nonalgorithmic, and between information processing and interactive systems. Even in the short distance travelled in this book, associative learning is seen to embrace both sides of these dichotomies.



Contents of Book



0    Preamble
1    Chapter 1    Associative Learning
2    Chapter 2    The BunPie Microworld
3    Chapter 3    Designing Templates
4    Chapter 4    Numbers in the Head
5    Chapter 5    Universal Turing Machine
6    Chapter 6    Communicating Intentions
7    Chapter 7    Consciousness before Language
8    Chapter 8    An Hierarchical Task
9    Chapter 9    Stress and Disapproval
10   Chapter 10   Painted Vision
11   Chapter 11   Cooperation
12   Chapter 12   Turn-Taking
13   Chapter 13   Climbing a Tree or Building a Rocket

R   References
A   Appendix A. Implementation of PURR-PUSS
             B. A Brief Historical Note about STeLLA
I   Index


Overview of Chapter 1

Associative learning is defined in terms of associations and association templates. The definitions are justified in terms of simple neural networks. The basics of associative learning are illustrated in a robot brain, called PURR-PUSS (PP). The structure of PP is justified in terms of a finite-state automaton with short-term and long-term memories.
PP has sufficient computational power and acquires its own top level program as a growing collection of associations. The system avoids the limitations of formal systems by close interaction with an open world.
This book aims to convince the reader that associative learning is a promising explanation of brain processing, in spite of authoritative claims to the contrary.

Overview of Chapter 2

The PURR-PUSS (PP) system is used to suggest how associative learning might be exploited to develop a robot intelligence. Four claims are made for PP: It becomes an individual through learning, it sets its own goals, it is computationally powerful, and it escapes the limitations of formal systems.
The learning of associations, motivation through novelty and reward, and the construction of plans are explained in terms of PP's actions and stimuli in the BunPie microworld.

Overview of Chapter 3

This chapter is a miscellany of topics, which will be referred to later in the book. It is suggested that you skip the chapter on a first reading and return to it as necessary.
Event-sustainability provides conditions for a set of templates to be used, particularly in making plans.
Generalization and discrimination are shown to be fundamental consequences of some associations having larger contexts than others.
The concept of an event in a context is extended to include threading events, predictions and even whole associations. New types of association template use these event-types and also include the important replacement association. These templates make possible different ways of holding and changing information.

Overview of Chapter 4

In this experiment associations combine to support the behaviour of a robot in a simple microworld.
When the real stimuli from the microworld disappear, predicted stimuli take over in the formation of associations. This allows the behaviour to become internalized.

Overview of Chapter 5

The aim of this chapter is to establish formally the computational power of associative learning by showing that a Universal Turing Machine can be learned. The cut down PP system learns the whole of the UTM including its extendable tape. Two special features of the demonstration are the five parallel streams of events and the teaching in parts. The demonstration illustrates the interweaving of associations to support the parallel streams. The teaching in parts enables the whole system to be taught before letting it emulate the UTM entirely on its own.

Overview of Chapter 6

In her book "The Emergence of Symbols", Elizabeth Bates (1979) describes a paradigmatic experiment to demonstrate what she refers to as "the First Moment" in the dawn of language. The experiment is so beautifully simple that I was tempted to try it out with PURR-PUSS. Not only was there here a well-defined experiment but also a careful analysis of the results. Furthermore, Elizabeth Bates saw this as the first of "two critical moments in the dawning of human communication through symbols". What better place could there be to start on the hazardous journey of giving PP language!
The plans of PP play a central role in the experiment of this chapter. Unrealized plans led to a build-up of Frustration until PP cried. The plans revealed what PP wanted to happen and we could check the plans against PP's memory at the time. This is a luxury that psychologists studying infants do not have.

Overview of Chapter 7

A future buyer of a conscious intelligent robot who wants the robot to carry out a task on its own must convince herself of the robot's competence and intentions.
A series of increasingly demanding definitions of robot behaviour are constructed with the needs of the robot buyer in mind. The series goes far beyond current progress with the PP system and indicates a route for future research.

Overview of Chapter 8

The 5-puzzle experiment shows how an auxiliary action, called `RaiseEyebrows', could be used to process a task with nested subtasks. The processing of nested subtasks is analogous to the processing of nested clauses in language or nested subroutines in programming.
Each time a subtask is entered, the name of the subtask is associated with the level of the subtask and the auxiliary action. At the end of the subtask (even if lower subtasks have been processed in the meantime), the name of the subtask is recovered and the subtask is continued.
Using soft teaching, the teacher needs to provide the auxiliary action just once for each task level, the first time that that task level is entered. After that, PP chooses the auxiliary action itself.
The next chapter demonstrates a much more acceptable auxiliary action for teacher to use when teaching nested clauses. This auxiliary action is the addition of stress to sounds, which is a common feature of ordinary speech.

Overview of Chapter 9

The aim of this chapter is twofold.
First, a small memory task is employed to show how stress on sounds can be used as an auxiliary action in soft teaching. This auxiliary action would be more appropriate for teaching language structure than the RaiseEyebrows auxiliary action used for the 5-puzzle task in the last chapter.
Secondly, the same memory task is used to demonstrate how disapproval operates to distinguish those associations responsible for an unwanted action from other associations active at the same time. In brief, disapproval of an action of PP causes all associations active at the time to be marked. When PP then goes through the same situation with the correct action and no disapproval, the still active (and therefore wanted) associations lose their disapproval marks.

Overview of Chapter 10

PP is given a special memory, called Painted Vision, which holds what it has recently seen in the space around it. The memory decays with time and is refreshed by PP's visual stimuli.
This chapter describes this memory and how it works. Then the stimulus-predicting templates are given for PP in the BunPie Microworld, so the contribution of Painted Vision can be emphasised. Finally, saccades are introduced to induce refreshing of the Painted Vision when something moves in peripheral vision.
Painted Vision is seen as a spatial equivalent of STM.

Overview of Chapter 11

Here we see the final runs of the BunPie Microworld with the two PP robots learning to cooperate. In the smallest 2x2 Microworld, the two PP robots lock into a cycle of perpetual cooperation some number of steps from the start. On different runs, different cycles are entered after different numbers of steps. If the robots were disturbed from one of these cycles they would be no better at cooperating than they were when they entered the cycle, so the situation seen in the 2x3 world, where there is a gradual increase in cooperative resources but no locked in cycle, is more robust and more deserving of the description `emergent behaviour'.
Until the robots enter a cycle and become like automata, they are acting on their own and are driven largely by their own novelty goals. We see how the PP robot can escape the limits of a formal system.
In the last part of the chapter, the two robots are separated by a wall and it is shown how cooperation can be supported by communication between the robots: each robot makes sounds and hears those made by the other robot.

Overview of Chapter 12

One robot is taught a simple counting task using ten fingers. It is then shown that another robot having the same association templates can learn this task from the first robot. A third robot has one of its templates lacking an event-type. When we let this third robot interact with the first robot, it can learn only part of the counting task.
With one robot learning from another it is important that they take turns in a disciplined way. To make this possible, the robots are provided with a turn-taking reflex controlled by a turn-taking frustration level.
The aim of the experiment is to illustrate the (obvious?) fact that robots can learn best from each other if they each have the same set of association templates.
The final section demonstrates a mechanism by which one robot can learn from imitating another.

Final Paragraph of Chapter 13

Dissatisfaction is a driving force for innovation and research. The collection of experiments described in this book is incomplete in every way. It is hoped that their incompleteness will draw others into this research area. The taboo on associative learning research must be lifted. It owes nothing to behaviourism. You have seen a little of what associative learning can do. Contrary to Pinker (1997), quoted at the beginning of Chapter 5, to Fodor (1994), quoted at the beginning of Chapter 8, and to Jackendoff (1992), quoted at the beginning of this chapter, the process of association is neither underpowered nor hopeless nor inadequate:
   Associative learning is potent and promising!


 

John Andreae can be contacted at: John Andreae

All cat line drawings © Gillian M Andreae