The role of simulations in Cognitive Science

Presented at The Science of Consciousness conferece in Taormina, may 2023

Transcript

1. UNIVERSITY OF GOTHENBURG The role of simulations in Cognitive Science ANDREAS CHATZOPOULOS, DEPARTMENT OF APPLIED INFORMATION TECHNOLOGY
2. UNIVERSITY OF GOTHENBURG SCIENTIFIC EXPLANATION OF CONSCIOUSNESS
3. UNIVERSITY OF GOTHENBURG § Hard to bridge explanatory gap with physical bridge laws § We should embrace biologically inspired models § We should work with simulations created from these models § How do we do this? Scientific explanation of consciousness
4. UNIVERSITY OF GOTHENBURG SCIENTIFIC EXPLANATION?
5. UNIVERSITY OF GOTHENBURG Starting point: Ongoing debate about explanations
6. UNIVERSITY OF GOTHENBURG MECHANISTIC EXPLANATIONS § Explains the behavior of a complex system with a model of the mechanistic interaction between its parts § Bottom-up approach that can capture emergent phenomena DYNAMICAL EXPLANATIONS § Uses dynamical models to study qualitative features regardless of component details § Describes how a system evolves over time Ongoing debate: Types of explanations
7. UNIVERSITY OF GOTHENBURG MECHANISTIC EXPLANATION A description of how the components (ion channels) are organized and interact to generate action potentials DYNAMICAL EXPLANATION Described with a set of coupled differential equations that illustrates the dynamics of the membrane potential Example: The action potential
8. UNIVERSITY OF GOTHENBURG SIMULATIONS
9. UNIVERSITY OF GOTHENBURG Models and simulations Reality Model Simulation representation of ”live” version
10. UNIVERSITY OF GOTHENBURG DISCRETE SIMULATIONS § Based on discrete space-time structure § Set of possible states of the system are assumed to be discrete CONTINUOUS SIMULATIONS § Underlying space-time structure and possible states are continuous § Formulated in differential equations Simulations: Different types Hartmann (2014) – “The world as a process: Simulations in natural and social sciences”.
11. UNIVERSITY OF GOTHENBURG Simulations and explanations Mechanistic explanation model Discrete simulation Dynamical explanation model Continuous simulation How shall we construct these?
12. UNIVERSITY OF GOTHENBURG Agent Based Modeling and Simulation (ABMS) § Paradigm for simulating actions and interactions of autonomous agents § Uses simulated agents for producing a macro level phenomenon § Makes it possible to discover hidden patterns – can highlight behaviors that are not derivable from the properties of the constituents How shall we construct discrete simulations?
13. UNIVERSITY OF GOTHENBURG ABMS example Claes Strannegård: ECOTWIN § Agent-based simulation of a simple ecosystem with hares and foxes § Each agent is driven by internal algorithms that simulates individual behavior
14. UNIVERSITY OF GOTHENBURG § Most often used in social science, biology and ecology § Proposal: We can also use this as a paradigm for modeling/simulating other mechanistic explanations in cognitive science § Good fit for mechanistic explanations - based on discrete components and their interactions ABMS for mechanistic explanations
15. UNIVERSITY OF GOTHENBURG Simulations and explanations Mechanistic explanation model ABMS Dynamical explanation model Continuous simulation
16. UNIVERSITY OF GOTHENBURG LARGE LANGUAGE MODELS?
17. UNIVERSITY OF GOTHENBURG MECHANISTIC? § Captures causal relationship between word and concepts DYNAMICAL § Learns the statistical relationship between words and concepts LLMs – Hard to categorize § Do not explicitly model the physical brain § The do seem to tell us something though § How shall we think about this?
18. UNIVERSITY OF GOTHENBURG HOW-ACTUAL EXPLANATION (HAE) § Illustrates a phenomena in exactly the way it occurs § A model of how things actually are HOW-POSSIBLE EXPLANATION (HPE) § Propositional model of how a phenomena might possibly occur § A model of how things could possibly be Philosophical analysis
19. UNIVERSITY OF GOTHENBURG HOW-ROUGLY EXPLANATION (HRE) § Less strict similarity requirement § Represents a target, roughly § Example: Gelder’s dynamical systems approach – Continuous interaction rather than discrete – Makes no claim to be a realistic representation of the brain – Goal is to replicate certain features in a better way than computational models do § In the same way, LLMs could also be considered HREs Philosophical analysis
20. UNIVERSITY OF GOTHENBURG § Both mechanistic and dynamical models can be turned into simulations § LLMs cannot 100% be categorized as mechanistic or dynamical § Philosophically, LLMs are HREs that can tell us important things about cognitive processes even though they are not explicit representations of the brain Conclusions
21. UNIVERSITY OF GOTHENBURG THANK YOU andreas.chatzopoulos@ait.gu.se @mimavox@zirk.us