Circuit sharing and the implementation of intelligent systems

Circuit sharing and the implementation of intelligent systems Michael L. Anderson Department of Psychology Franklin & Marshall College Lancaster, PA USA Institute for Advanced Computer Studies Program in Neuroscience and Cognitive Science University of Maryland College Park, MD USA 11/2008 AAAI 2008 1

Cognitive Architecture What is the overall functional architecture of the brain? 11/2008 AAAI 2008 2

Cognitive Architecture The classical, and still most widely accepted answer: 1. Low-level localization of function 2. High-level localization of domain 11/2008 AAAI 2008 3

Low-level localization of function Penfield’s Homunculus 11/2008 AAAI 2008 4

High-level localization of domain Brodmann map showing functional domains 11/2008 AAAI 2008 5

More abstractly 3 1 5 2 6 4 Classical c. a. (modularity? ) suggests: • Each brain area has a fixed working • Each function (and class of functions) is implemented in dedicated neural structures 11/2008 AAAI 2008 6

As opposed to 3 1 5 2 6 4 Holism (connectionist c. a. ? ) suggests: • Each brain area has a flexible working • Each function (or class of functions) is implemented in overlapping neural structures 11/2008 AAAI 2008 7

As opposed to 3 1 5 2 6 4 Redeployment suggests: • Each brain area has a fixed working • Each function (or class of functions) is implemented in overlapping neural structures 11/2008 AAAI 2008 8

What’s redeployment? Evolutionary considerations favor a “component re-use” model. Components evolved for one cognitive function are “exapted” for later uses. However, the original functionality is not lost—hence “redeployment” rather than exaptation. 11/2008 AAAI 2008 9

Evolution via redeployment 11/2008 AAAI 2008 10

Modularity vs. Holism vs. Redeployment 3 3 1 5 2 6 4 4 3 1 5 2 6 4 11/2008 AAAI 2008 11

Empirical evidence • Database of 665 (subtraction-based) imaging experiments in 20 cognitive domains. • “Functional connectivity” analysis of 472 experiments in 8 cognitive domains (all domains with > 30 experiments). 11/2008 AAAI 2008 12

Functional connectivity 1) Choose a spatial segmentation of the brain (we currently use Brodmann areas) 2) Choose an independent variable of interest (cognitive domain) 3) Determine which regions are statistically likely to be co-active, for different levels of the I. V. 11/2008 AAAI 2008 13

Step 3 in more detail A. Calculate chance probability (Q) of coactivation for each BA pair B. In each domain, determine observed probability (K) of co-activation of each BA pair C. Where there is a significant difference between Q and K (Χ 2), this is considered a “functional connection” 11/2008 AAAI 2008 14

Functional cooperation • Functional connection indicates areas that cooperate in service of cognition AB 11/2008 -AB Domain Co-active in domain -Domain Co-active not Not co-active in domain not in domain AAAI 2008 Not co-active in domain 15

List of domains Domain N Action 56 Attention 77 Emotion 42 Language 165 Memory 88 Mental imagery 31 Reasoning 33 Visual perception 57 11/2008 AAAI 2008 16

Functional cooperation Expected Coact. Prob Observed Coact. Prob Active. Area Co. Active. Area Chi. Square BA 10 L BA 32 L 0. 019 0. 036 8. 34 BA 10 L BA 32 R 0. 015 0. 054 61. 30 BA 10 L BA 40 L 0. 029 0. 054 13. 77 BA 10 L BA 40 R 0. 016 0. 036 13. 82 BA 10 L BA 44 L 0. 018 0. 036 10. 77 BA 10 L BA 44 R 0. 012 0. 036 28. 86 We can make graphs of these cooperation links. 11/2008 AAAI 2008 17

Action 18

Attention 19

Language 20

Comparing Domain Complexes Can compare many things, for instance: – Node overlap • Indicates B. A. s shared by different domain complexes – Edge overlap • Indicates functional connectivity/cooperation shared by different domain complexes – Network topology • May give clues about nature of function implementation 11/2008 AAAI 2008 21

Node vs. Edge Overlap Use Dice’s coefficient: 2(o 1, 2)/(n 1+n 2) Predictions: – Modularity: e, n – Holism: E, N – Redeployment: e, N 11/2008 AAAI 2008 22

Modularity vs. Holism vs. Redeployment 3 3 1 5 2 6 4 4 3 1 5 2 6 4 11/2008 AAAI 2008 23

Nodes vs. Edges 0. 9 Modularity Prediction 0. 8 Mean Dice's Coefficient 0. 7 0. 6 0. 5 Edge overlap 0. 4 Node overlap 0. 3 0. 2 0. 1 0 11/2008 Edge vs. Node overlap AAAI 2008 24

Nodes vs. Edges 0. 9 Holism Prediction 0. 8 Mean Dice's Coefficient 0. 7 0. 6 0. 5 Edge overlap 0. 4 Node overlap 0. 3 0. 2 0. 1 0 11/2008 Edge vs. Node overlap AAAI 2008 25

Nodes vs. Edges 0. 9 Redeployment Prediction 0. 8 Mean Dice's Coefficient 0. 7 0. 6 0. 5 Edge overlap 0. 4 Node overlap 0. 3 0. 2 0. 1 0 11/2008 Edge vs. Node overlap AAAI 2008 26

Nodes vs. Edges 0. 9 Actual Results 0. 8 Mean Dice's Coefficient 0. 7 0. 6 0. 5 Edge overlap 0. 4 Node overlap 0. 3 0. 2 0. 1 0 Edge vs. Node overlap p << 0. 001 11/2008 AAAI 2008 27

But. . . Maybe this result is just an artifact • Given a small number of nodes (84) • Large number of possible edges (3486) • Get high node overlap and low edge overlap just by chance 11/2008 AAAI 2008 28

0. 9 Results vs. Chance 0. 8 Mean Dice's Coefficient 0. 7 0. 6 Edge overlap 0. 5 Edge chance Node overlap 0. 4 Node chance 0. 3 0. 2 0. 1 0 Edge vs. Node overlap p << 0. 001 11/2008 AAAI 2008 29

4 implications 1. Give up on modularity in its classic form 2. Need to develop a domain-neutral vocabulary for cognitive science 3. Assigning computational/cognitive roles to brain areas will require cross-domain modeling 4. Should consider cross-domain uses when designing cognitive components 11/2008 AAAI 2008 30

4 implications 1. Give up on modularity in its classic form 2. Need to develop a domain-neutral vocabulary for cognitive science 3. Assigning computational/cognitive roles to brain areas will require cross-domain modeling 4. Should consider cross-domain uses when designing cognitive components 11/2008 AAAI 2008 31

Divergence in implementation Complex system Module 1 Component 1 Subcomponent 1 Component 2 Module 2 Component 3 Component 4 Module 3 Component 5 Component 6 . . . Modular architectures support functional assignment by decomposition and analysis 11/2008 AAAI 2008 32

Convergence in implementation Complex system Functional Complex 2 Functional Complex 1 Component 2 Subcomponent 1 Subcomponent 2 Component 3 Component 4 Subcomponent 3 Functional Complex 3 Component 5 Subcomponent 4 Component 6 Subcomponent 5 Sub-component 6. . . 11/2008 AAAI 2008 33

Cross-domain modeling 1. Cannot determine what a sub-component should do by considering only an individual task or task category, as been the normal practice. 2. Must begin to consider at design time the use of low-level components across multiple tasks in multiple domains. 11/2008 AAAI 2008 34

Cross-domain modeling (2) To do this: 1. Model each function of the system 2. Map sub-functions to a limited set of components 3. Constraint: each point of overlap must assign same (abstract) sub-function to each component 11/2008 AAAI 2008 35

Cross-domain modeling (3) 11/2008 AAAI 2008 36

• Anderson, M. L. (2007). The massive redeployment hypothesis and the functional topography of the brain. Philosophical Psychology, 21(2): 143 -174. • Anderson, M. L. (2007). Evolution of cognitive function via redeployment of brain areas. The Neuroscientist, 13(1): 13 -21. • Anderson, M. L. (2007). Massive redeployment, exaptation, and the functional integration of cognitive operations. Synthese 159(3): 329 -45. • Anderson, M. L. (2008). Circuit sharing and the implementation of intelligent systems. Connection Science, 20(4): 239 -51. http: //www. agcognition. org 11/2008 AAAI 2008 37