PhD thesis available: Neural mechanisms for control in complex cognition

[Patrick Simen]

Dear Connectionists,

My PhD thesis, 'Neural mechanisms for control in complex cognition', is 
now available at 

http://www.math.princeton.edu/~psimen/SimenThesis.pdf. 

The abstract and table of contents are presented below. I hope it will be 
of interest to you!

--Patrick Simen



ABSTRACT

Neural network models of complex cognitive tasks are difficult to build. 
Most previous work has focused on the difficulty of using structured 
symbolic representations in neural networks. This thesis focuses on the 
problem of control. During problem solving, some form of control is 
necessary for sequencing operations, for selecting actions, and for 
manipulating goal representations. I present a set of control mechanisms 
inspired and constrained by brain organization that are powerful enough to 
guarantee basic problem solving ability; in fact, I show that they are 
computationally universal. These mechanisms exploit a simple method for 
controlling the temporal characteristics of activation in continuous-time 
neural networks that makes neural control of complex processes possible in 
properly organized neural cognitive models. The basic computational 
primitive is inspired by corticostriatal loops in which the cortical 
component is composed of columns organized in layers. An input layer and 
an output layer each form winner-take-all networks. These layers are 
connected via a corticostriatal loop that produces a controllable amount 
of internal propagation delay in signal transmission from input layer to 
output layer. Modules can be composed hierarchically to produce 
goal-directed control circuits for cognitive models that are formally 
equivalent to finite automata and share many properties of symbolic 
production systems. These control circuits are instantiated in a neural 
cognitive model of the Tower of London problem-solving task. The model 
implements the assumption that dorsolateral prefrontal cortex is 
preferentially involved in representing subgoal information during problem 
solving, and that frontostriatal loop circuits provide a timing function 
that is critical for proper problem solving performance. Normal subject 
performance is accurately simulated by the model, and performance under 
conditions of simulated prefrontal lesions and Parkinson's disease 
captures speed and accuracy impairments exhibited in patient data from the 
literature.



TABLE OF CONTENTS

1. Computational models of control
   1.1 Objective
   1.2 Defining control
       1.2.1 Control systems theory
       1.2.2 Control in formal computational systems
       1.2.3 Control in the brain
   1.3 Cognitive architectures based on production systems
       1.3.1 Working memory, goals and productions
       1.3.2 Conflict and resolution
       1.3.3 Learning
       1.3.4 Distributed control
   1.4 Existing neural models of control and symbolic processing
       1.4.1 Models of neural symbol processing
       1.4.2 Models of control
   1.5 A neural cognitive architecture
   1.6 Summary

2. Controlling and exploiting the temporal dynamics of neural activation
   2.1 Neural activation and positive feedback
       2.1.1 The activation function
       2.1.2 Self-excitation
   2.2 Measuring and encoding duration
   2.3 Summary

3. Using control to implement computationally universal neural primitives
   3.1 Finite automata and Turing machines
       3.1.1 Finite automata as control devices
       3.1.2 Turing machines
   3.2 Neural finite automata and Turing machines
   3.3 Components of continuous-time neural finite automata and Turing 
machines
       3.3.1 Encoding internal state
       3.3.2 Representing discrete values
       3.3.3 Implementing voltage sources
       3.3.4 Implementing simple logic functions
       3.3.5 Implementing simple memory devices
       3.3.6 Implementing gates and flip-flops
       3.3.7 Delay in columnar networks
       3.3.8 Clocks
   3.4 Implementing finite automata
       3.4.1 Maintaining internal state and encoding acceptance
       3.4.2 Input formats
       3.4.3 Computing the next state
   3.5 Neural tape mechanisms
   3.6 Summary

4. Using control to implement simplified neural production systems
   4.1 Production systems, classifier systems and control
   4.2 Defining neural productions
       4.2.1 Productions are connections between modules
       4.2.2 Productions are atomic and require effective representation
       4.2.3 Limitations of the production-connection mapping
   4.3 Activation regulation for conflict resolution
       4.3.1 Preferences
       4.3.2 Safe ramp-up rates in regulators
       4.3.3 When voting stops
       4.3.4 Combining excitatory and inhibitory regulators
   4.4 Goals
   4.5 Impasse detection
   4.6 Summary

5. A neural model of the Tower of London task
   5.1 Basic model structure
   5.2 Tower of London model
       5.2.1 Sensorimotor backbone
       5.2.2 Perceptual reasoning
       5.2.3 Move selection and gating
       5.2.4 Goals
       5.2.5 Subgoals
       5.2.6 Implemented algorithm
   5.3 Neural convergence detection and subgoal generation
   5.4 Performance of the model
   5.5 Summary

6. Mapping the computational architecture onto cortex and corticostriatal 
loop circuits
   6.1 Modules map onto cortex
       6.1.1 Laminar structure
       6.1.2 Columnar structure
       6.1.3 Regional mapping
   6.2 Column structures map onto corticostriatal loop circuits combined 
with cortical columns
       6.2.1 More detailed circuitry
       6.2.2 Cognitive functions and their impairments by disease
       6.2.3 Corticostriatal analogues in the column primitive
       6.2.4 Discussion
   6.3 Activation regulators map onto anterior cingulate cortex
   6.4 Summary 

7. Simulating the behavior of normal controls, prefrontal patients and 
Parkinson's patients on the Tower of London task
   7.1 Predictions of DLPFC mapping
   7.2 Predictions of frontostriatal mapping
   7.3 Summary

8. Discussion
   8.1 Summary
   8.2 Contributions
       8.2.1 A focus on control
       8.2.2 An hypothesis regarding brain organization and psychological 
function
       8.2.3 A means for temporal coding in neural networks
       8.2.4 Demonstrates flexible control through finite automaton and 
Turing machine emulation
       8.2.5 An example of the power of symbolic dynamics
       8.2.6 A simple method for the construction of complex neural 
cognitive models
       8.2.7 Mechanisms that use analog quantities for computation
       8.2.8 Demonstrates the value of committing to a low-level physical 
model of neural processing
   8.3 Remaining issues
       8.3.1 Synaptic modification
       8.3.2 The binding problem
       8.3.3 More realistic neurons


Appendix A. Sequence and duration learning
   A.1 Introduction
   A.2 Predictive error driven learning
   A.3 Computational motivations for laminar structure
       A.3.1 Asymmetric connection learning in recurrent networks
       A.3.2 Inhibitors of input and output
       A.3.3 A synaptic triad mechanism for learning
       A.3.4 Recruit-driven timing mechanisms for plasticity
       A.3.5 Recruitment of columns
   A.4 Learning durations with inhibitory strength modification
       A.4.1 Preventing propagation during recording
       A.4.2 The rate of weakening
       A.4.3 Isolating precision components from fluctuations
   A.5 Performance of the full sequence learning circuit

Appendix B. Glitches and glitch protection

Appendix C. Goals and a goal stack mechanism
   C.1 Properties of goals
   C.2 Stacking goals
   C.3 Performance of the goal stack
   C.4 Incorporating the goal stack mechanism into a column
   C.5 The goal stack as a tool for cognitive modeling




*************************************************************************
Patrick Simen                           209 Fine Hall                
Research Fellow                         Washington Rd. 
Center for the Study of Brain,          Princeton, NJ 08544-1000
    Mind and Behavior
Program in Applied and                  Phone: (609) 258-6155
    and Computational Mathematics       Fax:   (609) 258-1367
Princeton University                    email: psimen at math.princeton.edu
*************************************************************************