Course outline
Module 1
Module 1: Foundations of Cognitive Science
Foundations of Cognitive Science
Knowledge check
Q1. Which field is generally considered to have contributed the study of the mind as an information-processing system in the mid-20th century, alongside psychology, linguistics, and computer science?
- A. Astronomy
- B. Cognitive science ✓
- C. Geology
- D. Botany
Why: Cognitive science emerged in the 1950s and 1960s as an interdisciplinary field that combined psychology, linguistics, computer science, philosophy, and neuroscience to study mental processes as information processing.
Q2. What is the term for the view that mental states can be understood as computational processes operating on symbols?
- A. Behaviorism
- B. Connectionism
- C. Computationalism ✓
- D. Phenomenology
Why: Computationalism holds that cognition is a form of computation, where mental representations are manipulated according to rules, a foundational idea in both cognitive science and classical AI.
Q3. Which researcher proposed the Turing Test as a way to evaluate whether a machine can exhibit intelligent behavior indistinguishable from a human?
- A. John Searle
- B. Noam Chomsky
- C. Alan Turing ✓
- D. Herbert Simon
Why: Alan Turing proposed the Turing Test in his 1950 paper 'Computing Machinery and Intelligence' as an operational criterion for machine intelligence based on indistinguishable conversational behavior.
Q4. In cognitive science, what does the term 'representation' generally refer to?
- A. A physical copy of an object in the brain's anatomy
- B. An internal state or structure that stands for something in the world ✓
- C. A person's public statements about their beliefs
- D. The speed at which neurons fire
Why: A mental representation is an internal state or structure that carries information about, or stands for, something external, such as an object, concept, or proposition. Representations are central to explaining how minds process information.
Q5. Which level of analysis, proposed by David Marr, concerns the abstract problem a cognitive system solves and the logic of the solution?
- A. Implementation level
- B. Algorithmic level
- C. Computational level ✓
- D. Behavioral level
Why: Marr's computational level specifies what problem a system solves and why, independent of how it is carried out. The algorithmic level describes the processes and representations used, and the implementation level describes the physical realization.
Module 2
Module 2: Neural Networks and the Brain
Neural Networks and the Brain
Knowledge check
Q1. Which biological structure inspired the basic unit of an artificial neural network?
- A. The neuron ✓
- B. The synapse
- C. The axon hillock
- D. The glial cell
Why: Artificial neurons (nodes) were loosely modeled on biological neurons, which receive inputs, integrate them, and produce an output signal.
Q2. In an artificial neuron, what do the weights primarily represent?
- A. The number of layers in the network
- B. The strength of the connection between inputs and the neuron ✓
- C. The speed at which the neuron fires
- D. The total number of neurons in the brain
Why: Weights scale each input's influence on the neuron, analogous to how synaptic strength determines how much one neuron affects another.
Q3. What biological process is most closely analogous to adjusting weights during neural network training?
- A. Myelination of axons
- B. Synaptic plasticity ✓
- C. Apoptosis of neurons
- D. Neurotransmitter breakdown
Why: Synaptic plasticity, such as Hebbian strengthening or weakening of synapses, parallels how learning algorithms adjust connection weights based on experience or error.
Q4. What is the role of an activation function in an artificial neuron?
- A. It stores the training data
- B. It introduces nonlinearity so the network can model complex patterns ✓
- C. It randomly deletes connections to prevent overfitting
- D. It counts the number of inputs received
Why: Without nonlinear activation functions, stacked layers would collapse into a single linear transformation; nonlinearity lets networks represent complex relationships, loosely like the threshold behavior of real neurons.
Q5. Which statement best describes a key difference between biological brains and typical artificial neural networks?
- A. Artificial networks use electrochemical signals, while brains use digital signals
- B. Brains are far more energy-efficient and operate with asynchronous, spiking activity, unlike most artificial networks ✓
- C. Artificial networks have billions of neurons, while brains have only a few thousand
- D. Brains cannot learn from experience, while artificial networks can
Why: The human brain runs on roughly 20 watts using sparse, asynchronous spiking, whereas standard artificial networks use synchronous numerical computation that demands far more energy.
Module 3
Module 3: Symbolic vs Connectionist Approaches
Symbolic vs Connectionist Approaches
Knowledge check
Q1. Which approach to AI represents knowledge primarily as explicit symbols and rules that are manipulated by logical operations?
- A. Connectionism
- B. Symbolic AI ✓
- C. Reinforcement learning
- D. Evolutionary computation
Why: Symbolic AI, also called GOFAI, represents knowledge as explicit symbols and rules (such as in expert systems and logic-based reasoning) that are manipulated according to formal procedures. Connectionism instead stores knowledge in the weights of networks of simple interconnected units.
Q2. What is a key characteristic of connectionist models of cognition?
- A. Knowledge is stored in explicit, human-readable rules
- B. Knowledge is distributed across weighted connections between simple processing units ✓
- C. Each concept is stored in exactly one dedicated symbol
- D. Processing relies entirely on predefined logical proofs
Why: Connectionist models, inspired loosely by neural networks, distribute knowledge across many weighted connections between simple units. Learning occurs by adjusting these weights, so no single unit typically holds a specific concept.
Q3. Which of the following is a well-known limitation of purely symbolic systems that connectionist approaches were partly developed to address?
- A. They cannot perform any arithmetic
- B. They struggle with pattern recognition and learning from noisy, perceptual data ✓
- C. They require massive parallel hardware
- D. They cannot represent logical negation
Why: Symbolic systems handle explicit rule-based reasoning well but often struggle with perceptual tasks like vision and speech, where inputs are noisy and rules are hard to write by hand. Connectionist models excel at learning such patterns from data.
Q4. The debate between symbolic and connectionist approaches in cognitive science is often framed around which central question?
- A. Whether computers can be faster than humans
- B. How mental representations are structured and how cognition is implemented in the mind or brain ✓
- C. Whether programming languages should be typed
- D. Which hardware architecture is most energy efficient
Why: The debate concerns how mental representations are structured (symbolic, discrete, compositional versus distributed, subsymbolic) and how cognitive processes are implemented. It bears directly on theories of human thought as well as AI design.
Q5. Which example best illustrates a symbolic AI approach?
- A. A neural network trained to classify handwritten digits
- B. An expert system that uses if-then rules to diagnose medical conditions ✓
- C. A deep convolutional network that recognizes faces
- D. A recurrent network that generates text by predicting the next word
Why: Expert systems encode domain knowledge as explicit if-then rules that an inference engine applies, which is the hallmark of symbolic AI. The other examples rely on learned weights in neural networks, which are connectionist.
Module 4
Module 4: Learning, Memory, and Representation
Learning, Memory, and Representation
Knowledge check
Q1. In both cognitive science and machine learning, what does the term 'representation' most generally refer to?
- A. The physical hardware that stores information in a brain or computer
- B. An internal encoding of information that captures relevant features of the world for use in reasoning or prediction
- C. The process of communicating results to other researchers or users
- D. A fixed set of rules that a system must always follow
Q2. Which memory concept in cognitive science is most closely paralleled by the weights of a neural network that are adjusted during training?
- A. Working memory, which holds a few items temporarily
- B. Long-term memory, where learned knowledge is stored in persistent form
- C. Sensory memory, which briefly holds raw input
- D. Metacognition, which monitors one's own thinking
Q3. What is 'catastrophic forgetting' in neural networks, and what human-memory analogy is often drawn?
- A. A network becoming too large to train, analogous to information overload in humans
- B. A network abruptly losing previously learned knowledge when trained on new data, analogous to new learning disrupting old memories
- C. A network failing to store any data at all, analogous to amnesia from injury
- D. A network forgetting its training objective, analogous to losing focus during a task
Q4. In supervised learning, what role does the 'label' play in a training example?
- A. It is the input the model receives to make a prediction
- B. It is the correct output the model is trained to predict, analogous to feedback a learner receives
- C. It is a random number added to prevent overfitting
- D. It is the name of the algorithm used for training
Q5. Which statement best describes the relationship between reinforcement learning and behavioral psychology?
- A. Reinforcement learning ignores reward signals entirely, unlike behaviorism
- B. Reinforcement learning shares core ideas with behaviorist conditioning, such as learning actions that lead to rewards and avoiding those that lead to punishment
- C. Reinforcement learning only applies to images and has no link to psychology
- D. Reinforcement learning was developed independently of any theories of learning in animals
Module 5
Module 5: Perception and Computer Vision
Perception and Computer Vision
Knowledge check
Q1. Which early computer vision idea was directly inspired by the discovery that neurons in the visual cortex respond to edges with specific orientations?
- A. Convolutional filters that detect local edge orientations ✓
- B. Recurrent memory cells that store sequences of pixels
- C. Attention mechanisms that weight distant image regions
- D. Generative adversarial training of image synthesizers
Why: Hubel and Wiesel found that simple cells in the visual cortex respond to oriented edges in small receptive fields. Convolutional filters in neural networks, such as those in the Neocognitron and CNNs, mimic this by detecting local oriented patterns across the image.
Q2. What is the 'receptive field' of a neuron in the visual system?
- A. The total number of neurons connected to the retina
- B. The region of the visual field where a stimulus affects that neuron's firing ✓
- C. The speed at which a neuron responds to light
- D. The color channel a neuron is tuned to detect
Why: A receptive field is the specific region of the visual field in which a stimulus will alter a given neuron's activity. Convolutional neural networks use a similar concept: each unit in a layer only sees a local patch of the input.
Q3. Why is the problem of inferring 3D structure from a 2D retinal image considered ill-posed?
- A. Retinal images have too few pixels to be useful
- B. Many different 3D scenes can produce the same 2D projection ✓
- C. Light cannot be recorded by biological or artificial sensors
- D. Depth information is always discarded by the eye
Why: Projection from 3D to 2D loses depth information, so a single image is consistent with infinitely many scenes. Perception and computer vision must rely on priors, multiple cues, or learned assumptions to resolve this ambiguity.
Q4. Which of the following is an example of a visual illusion that suggests the brain relies on prior expectations when perceiving images?
- A. The Necker cube, whose depth orientation flips spontaneously ✓
- B. Color blindness in people with missing cone cells
- C. Afterimages that appear after staring at a bright light
- D. Myopia, where distant objects appear blurry
Why: The Necker cube reverses its apparent 3D orientation even though the drawing never changes. This shows the perceptual system imposes an interpretation on ambiguous input, similar to how Bayesian and predictive models of vision explain perception as inference.
Q5. What was the main purpose of the ImageNet challenge in the development of modern computer vision?
- A. To measure the energy efficiency of image processors
- B. To provide a large labeled benchmark for training and comparing object recognition systems ✓
- C. To test whether machines can draw realistic paintings
- D. To compare the speed of different camera sensors
Why: ImageNet provided over a million labeled images across thousands of categories. The annual competition drove rapid progress, most notably the 2012 breakthrough by AlexNet, a deep convolutional network, which demonstrated the power of deep learning for visual recognition.
Module 6
Module 6: Language, Reasoning, and Large Models
Language, Reasoning, and Large Models
Knowledge check
Q1. What is the core training objective of most large language models?
- A. Classifying images into predefined categories
- B. Predicting the next token in a sequence given the preceding tokens ✓
- C. Maximizing reward through trial-and-error interaction with an environment
- D. Clustering documents by topic without any labels
Why: Most large language models are trained with next-token prediction, learning to estimate the probability of each token given the context that precedes it. Image classification, reinforcement learning, and clustering are different paradigms.
Q2. Which cognitive science concept is most closely related to the idea that language models learn meaning from patterns of word co-occurrence?
- A. Dual-process theory
- B. Distributional semantics ✓
- C. Working memory capacity
- D. Behaviorist conditioning
Why: Distributional semantics holds that a word's meaning can be inferred from the contexts in which it appears. This is the principle underlying word embeddings and, by extension, large language models.
Q3. What does the term 'tokenization' refer to in the context of large language models?
- A. Encrypting text so it can be stored securely
- B. Splitting input text into discrete units such as words, subwords, or characters that the model processes ✓
- C. Assigning a confidence score to each generated sentence
- D. Removing grammatical errors from training data
Why: Tokenization converts raw text into a sequence of discrete units called tokens, which are then mapped to numerical IDs and embeddings. Subword tokenization is common because it balances vocabulary size with coverage of rare words.
Q4. Chain-of-thought prompting is a technique where a model is encouraged to do what?
- A. Produce answers in a single word only
- B. Generate intermediate reasoning steps before giving a final answer ✓
- C. Ignore the prompt and rely on memorized responses
- D. Translate its internal weights into plain English
Why: Chain-of-thought prompting elicits step-by-step intermediate reasoning, which often improves performance on arithmetic, logic, and multi-step problems. It is an observable behavioral technique rather than a change to the model's architecture.
Q5. Why do researchers in cognitive science find the comparison between human language acquisition and LLM training both illuminating and contested?
- A. Because LLMs are trained on far less data than children receive
- B. Because LLMs learn from vastly more text than humans encounter, raising questions about whether similar mechanisms underlie both ✓
- C. Because humans cannot learn grammar without explicit instruction, unlike models
- D. Because both humans and LLMs rely exclusively on visual input
Why: LLMs are trained on orders of magnitude more text than a child hears, and they lack embodied, social experience. This makes the comparison useful for testing theories of language learning while remaining contested as to whether the underlying mechanisms are similar.
Module 7
Module 7: Ethics, Interpretability, and Future Directions
Ethics, Interpretability, and Future Directions
Knowledge check
Q1. Which of the following best describes the goal of interpretability research in AI?
- A. Making AI systems faster to train on large datasets
- B. Understanding the internal mechanisms by which a model arrives at its outputs
- C. Reducing the energy consumption of data centers
- D. Increasing the number of parameters in a neural network
Q2. Why do many researchers argue that cognitive science can inform the ethical design of AI systems?
- A. Cognitive science provides exact rules for moral behavior
- B. Cognitive science studies how humans reason, perceive, and make decisions, which helps identify where AI may diverge from or mimic human judgment
- C. Cognitive science has proven that machines cannot make mistakes
- D. Cognitive science replaces the need for legal and ethical oversight
Q3. What is the 'black box' problem in machine learning?
- A. Models are physically enclosed in secure hardware that cannot be accessed
- B. The internal reasoning of complex models is difficult for humans to inspect or explain
- C. Training data is always stored in an encrypted format
- D. Neural networks can only process one input at a time
Q4. Which concept from cognitive science is most closely related to the idea of humans being able to understand and trust AI explanations?
- A. Operant conditioning
- B. Mental models
- C. Circadian rhythms
- D. Neurotransmitter release
Q5. Which ethical concern is most directly associated with AI systems that learn from historical human data?
- A. The systems will become too slow to be useful
- B. The systems may reproduce or amplify biases present in that data
- C. The systems will require no human maintenance
- D. The systems will always outperform human experts in every domain