Exam Study Notes

The Structure of Thought

  • A mind is a computational system; computations are performed in cognition.
  • Classical Computational Theory of Mind:
    • The mind is a formal system.
    • Thoughts are logically constructed with mental language.
  • Formal System:
    • Symbols are combined into expressions and manipulated into processes.
    • Mental language and natural language must connect.
    • Properties of natural language can indicate properties of mental language.

Language of Thought (LOT)

  • Not only the format for mental representations but also includes things like mental maps and imagery.
  • Properties of LOT:
    • Mental symbols possess semantic properties.
    • They represent something.
Compositionality:
  • Complex expressions are built from simpler expressions (constituents).
  • The meaning of a complex expression is a function of the meaning of its constituents and how they are combined.
  • Smallest Constituents: Meaningful units (e.g., "John jumped," "Mary climbed")
Logical Structure:
  • Expressions have logical form; they contain analogues of logical operations (and, if-then, etc.).
  • Modeled using predicate calculus.
    • Example: "John jumped and Mary climbed" can be represented as: Jump(John)∧Climb(Mary)Jump(John) \land Climb(Mary)
Syntactic Processes:
  • Argument from Productivity:
    • We can entertain a potential infinity of thoughts.
  • Argument from Systematicity of Thought:
    • Thoughts are related in systematic ways.
    • Understanding one thought demonstrates understanding related thoughts.
      • Example: If you understand "Bluey gave a flower to Greeny," you understand "Greeny gave a flower to Bluey."
  • Argument from Systematicity of Thinking:
    • Systematic relations exist among inferences.
    • Cognitive/linguistic capacities must have similar concepts.
Language of Thought Hypothesis:
  • Thinking occurs in a mental language that is compositional, has logical form, and follows syntactic rules.
  • Supported by arguments from practice, productivity, and systematicity.
  • Productivity: Ability to entertain a potential infinity of thoughts, similar to how we can generate a potential infinity of sentences in natural language.
  • Systematicity: Interrelations between the thoughts we can have and the ways in which we can think.
  • Thoughts are made up of simple symbols that can be recombined.
  • Related things share symbols in common:
    • Example: mim(BL,GR)mim(BL, GR) and mim(PI,GO)mim(PI, GO)
  • Computations manipulate symbols based on syntactic properties, not semantic properties.
  • Meaning does not factor into the rules of computation; only form does.
    • Example: "If John jumped then Mary climbed" vs. "John didn't not jump"
How to Build a Thought?
  • Markov Chain (Andrey Markov):
    • A system describing a sequence of distinct states, where the probability of each transition depends only on the state that came before.
    • Example: Weather patterns (sunny, cloudy, rainy) with probabilities for transitions between states.

Is Markov Enough?

Problems with Markov Models:
  • Problem #1: Nonsense phrases like "Colorless green ideas sleep furiously" - Markov models assign zero probability but are syntactically well-formed.
  • Problem #2: Long-distance dependencies such as "either…or" and "if…then" require the model to remember previous states over long sequences.
From Chains to Trees:
  • Generative Phrase-Structure Rules:
    • Resolve Problem #1.
    • Example: The happy boy eats ice cream.
  • Recursion:
    • Resolves Problem #2.
    • A rule that embeds one instance of a symbol inside another instance of the same symbol.
  • Hierarchical Structure: Symbols are chunked into larger sections/symbols.
  • Phrase structure is a better fit for NL; perception breaks down objects into part-whole structures.

Mind as Network

Symbolic Success Stories

  • The mind is a symbol manipulator - a computer.
  • What the mind does can be modeled with a Turing machine.
  • Some aspects of thought are better captured by a phrase-structure grammar.

Reconstructing Cognition

  • Instead of thoughts, we just need neurons, memories, connections, images, and activation levels.
Reasoning Types
  • Taxonomic: Relationships between categories (e.g., Dog is a subset of Mammal).
  • Temporal: Reasoning about times, durations, and time intervals.
  • Qualitative: Reasoning about the direction of change in interrelated quantities.
    • Example: If the temperature of gas in a closed container increases, then pressure increases.

Connectionism

  • The brain is nothing like a Turing machine; we need a different architecture.
  • Properties:
    • Emergence: Properties arise from system components working together.
Hebbian Learning
  • Neurons that fire in synchrony strengthen their connections.
  • "Neurons that fire together, wire together."
Perceptron
  • First neural network.
  • Many networks can compute the same function; the perceptron seems closer to meeting the requirements of a functional explanation.
  • Failed to mathematically prove that perceptrons could learn all classes of computations, which led to the abandonment of Connectionism.

Artificial Neural Networks

  • Nodes: Input, hidden, output layers
  • Connections: Activation function, weights
  • Breakthrough in 1986: The backpropagation algorithm, which allows hidden layers to be adjusted to minimize error.
Types of Learning
  • Supervised Learning: Training data plus supervision (classification, regression).
  • Unsupervised Learning: No training (clustering, association, dimensionality reduction).
  • Reinforcement Learning: Exploration and rewards (uses principles from behaviorism).

Mind as Network (Part 2)

Connectionism

  • Mind is a network of nodes and weighted connections.
  • Parallel processing solves the 100-step constraint.
  • Connectionism can do complex computations with simple stats; symbolic vs. distributed representation.
Strengths:
  • Noisy input or unit destruction causes graceful degradation.
Weaknesses:
  • Unable to perform one-off learning.
  • Require huge numbers of parameters and massive datasets.
  • Difficulties generalizing performance from training sets.
  • Unclear how to capture unbounded recursion & systematic thought.

What's in Your Mind?

Cognitive Capacity

  • What are the basic operations available to the mind, and what is its capacity?
Cognitive Architecture
  • The basic operations available to the mind and the way it encodes, stores, and retrieves information.
Constraints on Behavior
  • Sources of behavioral regularities:
    1. What we know about the world (representations).
    2. The way the mind is wired up (architecture).
Memory
  • An active system that receives, stores, organizes, alters, and recovers information.
    • Encoding: Converting information into a usable form.
    • Storage: Holding information in memory for later use.
    • Retrieval: Taking memories out of storage.
Sensory Memory to Long Term Memory

SENSORY→SelectiveAttention→ShortTermMemory→LongTermMemorySENSORY \rightarrow SelectiveAttention \rightarrow ShortTermMemory \rightarrow LongTermMemory

  • Short Term Memory (STM) or Working Memory:
    • "Workbench" with limited storage capacity.
The Magic #7
  • Plus or minus 2
  • Channel capacity: Max information processed by the cognitive system.
  • Chunking: Grouping units together to remember more details; flexible slots in working memory.

Working Memory

  • Constructive operations, transformations (chunking, recoding), and decision processes (identification, comparison).

Memory Search

  • How is information organized in working memory, and what's the search algorithm?
Predictions
  • Parallel Search: Reaction time is the same regardless of set size.
  • Serial Search: Reaction time increases with set size.
    • Self-Terminating: Search stops once the object is found.
    • Exhaustive: Search goes to the end of the line even after finding the object.
Results
  • Channel capacity: Working memory is flexible.
  • Memory search: Default is serial + exhaustive.

Mental Modules

Modularity

  • The mind is made up of independent systems/modules specialized for certain computations.
Module characteristics:
  1. Fixed neural architecture.
  2. Domain specificity.
  3. Mandatory operation.
  4. Limited central accessibility.
  5. Informational encapsulation.
  • Prosopagnosia: Inability to recognize faces. (faceblindness)
  • Blindsight: No conscious visual experience due to visual cortex damage.

Vision

  • Rods and cones help with vision.
  • Sensation: Detection (bottom-up).
  • Perception: Understanding (top-down).
Processing Types
  • Bottom-Up: Information flows from senses to perception.
  • Top-Down: Information flowing from world knowledge to perception.

Perceptual Grouping

  • Encapsulation: Restriction of information flow into a system.
  • Cognitive Impenetrability: Processes of perception are impenetrable to information in central memory.
    • Examples given: backpack/hill, distance estimations, reflective darkness estimations
  • Pareidolia: The tendency to see ordered structures in visual patterns (faces in objects, animals in clouds).
Laws
  • Law of Proximity: Grouping based on arrangement.
  • Law of Similarity: Grouping holistic organization in the brain.
  • Gestalt psychology influences top down perceptions

Concepts and Schemas

Concepts

  • Symbolic: Concepts stand for an idea/object but doesn't have a genuine resemblance.
  • Compositional: Complex concepts built up from simple concepts.
Classical Theory
  • Concepts are complex representations that encode necessary and sufficient conditions.
Prototype Theory
  • Concepts are complex mental representations that encode conditions that items in their extension tend to have.
Strengths
  • Fuzziness
  • Typicality effects
Weaknesses

*Conceptual stability

Conceptual Atomism
  • Lexical concepts don't have any structure.
Theory Theory
  • embedded in mental structures

Schemas

  • Mental structure that helps us organize, interpret, and store information.
  • Assimilation: New information is modified to fit an existing schema.
  • Accommodation: An existing schema is modified, or a new one is created.
  • Object Schemas: Help us understand and interpret objects (e.g., house).
  • Scene Schemas: Help us interpret new situations (e.g., university department).
  • Causal Schemas: Guide reasoning about cause and effect (e.g., launching schema).
  • Event Schemas: Using schemas to remember details (e.g., child's birthday party).

The Social Mind

Evolutionary Approach

  • Humans are an evolved species.
  • Survival hinges upon cooperation.
Lineage:
  • Homo sapiens evolved into modern form 200,000 years ago.
  • Environment of evolutionary adaptation.

Eusociality

  • Humans are eusocial animals.
  • Humans live in complex hierarchal societies w/ lots of members & relationships.

Social Contract Theory

  • Humans evolved domain-specific cognitive processes (module) for reasoning about social contracts.
  • Cheater-Detection Algorithm: Procedures for figuring out when someone has violated a social contract.

The Challenge of Cooperation

  • Mechanism for the evolution of cooperation.
  • Reciprocity: Social norm of returning the favor.
    • Direct: Between the same two individuals.
    • Indirect: Known by a third party.
  • Reciprocity is the default.

Theory of Mind

  • We model the outside world via mental representations.
  • We also model the internal mental states of others.
  • The ability to model thoughts of thinking of how others see, think of us, & thinking of their opinions.

The Social Mind

Evolution of Intelligence
  • Cultural intelligence hypothesis: Social cognition preceded general intelligence.
Intelligence
  • Physical intelligence: Spatial reasoning, sense of causality, object permanence.
  • Social Intelligence: communication + theory of mind + Cooperation + joint attention

Distributed Cognition

  • Cognitive processes are not bound by the individual.
  • May involve coordination between internal and external structures.
  • May be distributed through time or across members of a social group.