lecture 1 - why modeling (and how)

Question 1

3 different disciplines

Answer 1

1. neuroscience
   - the physical universe
   - data-rich but theory poor
   - lots to measure, hard to interpret
2. cognitive science/psychology
   - the human mind
   - long-standing debates on theories
   - theories lack mechanistic explanations
3. artificial intelligence
   - aims to create intelligent systems that are possibly, but not necessarily, inspired by brains

Question 2

how to bridge gaps between disciplines

Answer 2

by using physical science for detailed, mechanistic theories of the mind and behavior

(models of the mind grounded in measurable physical processes)

Question 3

David Marr’s levels of analysis

Answer 3

1. computational
   - why
   - focuses on what abstract problem the system is trying to solve.
   - asks “why does this system exist?” and “what is the goal of the computation?”
   - goals, functions, principles
2. algorithmic
   - how
   - examines what rules/processes/algorithms are used to solve the problem.
   - could technically be written down in pseudocode
   - still very abstract
   - rules, operations, transformations
3. implementation
   - what
   - how does the brain implement it in ‘wetware’
   - asks, “how is this computation physically realized?”
   - hardware, physiology, mechanisms

Question 4

what is a model

Answer 4

• a simplified version of reality
• easier to create than a parallel reality
• keeps/highlights the fundamental mechanisms that make an aspect of reality work, while leaving out unnecessary details
• by forcing us to decide what to include or exclude, models push us to think deeply about the workings of a system

Question 5

Richard Feynman: “what I cannot create, i do not understand.”

Answer 5

• truly understanding something means being able to break it down, explain it simply, and even recreate it
• so, if we understand cognitive processes well, we should be able to replicate it with a computer using explicit rules
• this forces us to write explicit rules to capture and simulate aspects of cognition

Question 6

classical experimentation approach in psychology

Answer 6

• hypothesis testing
• H0 is proposed and then tested, with the goal of finding evidence to reject it in favor of H1

Question 7

problems with hypothesis testing

Answer 7

1. p-hacking
2. replicability crisis

Question 8

how to fix problems with hypothesis testing

Answer 8

1. focus on mechanisms that underlie behavior and cognition and care less about the words in our theories
2. precision in hypotheses: clearer, more specific hypotheses can help avoid ambiguity in testing, making results more interpretable and reliable
3. move from hypotheses to prediction based models: compare models in how they explain our (left-out) data. this way researchers can more objectively assess which model better captures the phenomenon

Question 9

benefits of usisng models in scientific research

Answer 9

1. integrative: models are conceptually broad – they converge findings from a wide range of research.
2. falsifiable: precisely defined preditions – in contrast to theories that allow for flexible interpretations, models make precise predictions that can be tested and potentially proven wrong
3. detailed/specific: models are functionally complete – a phenomenon can be understood if it can be explained in a program and it produces the to-be-explained behavior/patterns/data
4. generative: playground for exploring new ideas and concepts – new concepts, ideas, and understanding can emerge from simulations

Question 10

weaknesses of models

Answer 10

1. high threshold: developing and understanding models often requires significant prerequisite knowledge
2. it’s not a real field: modeling is more of an application (applied tool) than a standalone scientific field
3. mathematical: typically involve detailed mathematical formulations, which can be complex and challenging
4. few standard methods: lack of standardization in modeling practices – everybody has their own model and no one wants to look at the other researcher’s

Question 11

terminology & mechanistic understanding

Answer 11

we need to be critical of terminology: it needs to be able to be translated into mechanistic understanding, so that it can improve our understanding of mind/brain function

Question 12

types of models

Answer 12

1. data-driven models: more statistical – focus on analyzing and interpreting data patterns
2. process/generative models: more conceptual – focus on describing the underlying mechanisms or processes that produce observable phenomena

• this is not a fundamental divide. these modelling efforts can work together, and are often combined in practice

Question 13

old AI vs. after the silicon revolution

Answer 13

old AI
- used to make explicit generative models based on abstract ideas
- these models tried to capture high-level cognitive processes by explicitly defining the steps or rules involved
- these models were successful, but in many cases had a hard time reaching human-level performance even in simple tasks
- this highlights the challenges of explicitly coding intelligence

since the silicon revolution
- more and more use ‘deep’ learning: ANNs inspired by the brain
- these models are loosely inspired by the brain’s structure and focus on distributed computation rather than explicit, step-by-step modeling.
- deep learning doesn’t rely on predefined rules or abstract ideas.

this is a pendulum
- there are things that deep learning cannot do

Question 14

Allen Newell

Answer 14

asked ‘how can the human mind occur in the physical universe?’