CPSY 1950 Deep Learning in Brains, Minds & Machines

NeuroAI goals, course structure, and how we will simulate scientific conferences (lightning talks and projects).

📑 Google Slides

No readings or pre-class activities for Week 1—this is the introductory session.

Goal: Put everyone on a leveled playing field for Week 3. Complete the Required Core first (mandatory for all), then the conditional tracks that match your background. You do not need to do every track.

Note on access to papers: For any paper link below, please use the Brown Library proxy links (you may be prompted to log in with Brown credentials off campus).

Required Core — Mandatory for all students

⚠️ Complete both items below before moving on to any conditional tracks. Students with a strong background in cognitive neuroscience may skim, but should still review to ensure familiarity with the course’s framing and terminology.
Expected time: ~90 min total

1. Marr’s Levels of Analysis (video)
   Nancy Kanwisher: “How can we study the human mind and brain? Marr’s levels of analysis”

2. Coarse brain organization (video; focus on gross organization + functions)
   MIT OCW 9.13 (Spring 2019) Lecture 2: “Neuroanatomy”
   Viewing guidance: Watch for the high-level tour of major brain divisions/structures and what they do. Pay particular attention to concepts of receptive fields, neural tuning, and cortical maps — these are fundamental for understanding how neural networks relate to biological systems. You do not need to know fine anatomical details.

Conditional Track A — Cognitive neuroscience foundations

Complete this track if you have not taken at least 1–2 courses in neuroscience or cognitive neuroscience.
Target time: up to ~180 min.

• A1) Neurons, spikes, firing rate (high-level; no ion-channel deep dive)
  BrainFacts: “How neurons communicate”

• A2) Quick map of structures + functions
  NIH/NINDS: “Brain basics — know your brain”

• A3) Methods lectures (videos)
  • MIT OCW 9.13 Lecture 4: “Cognitive neuroscience methods I”
  • MIT OCW 9.13 Lecture 5: “Cognitive neuroscience methods II”

  Note: Don’t worry if the jargon feels overwhelming — you don’t need to memorize all the details right now, and you can always come back to these videos later when these concepts become needed. These lectures include a section on Marr’s levels of analysis; you can skip it if redundant with the required core, though reviewing it twice may be helpful if you haven’t taken cognitive science courses.

• A4) Optional “what measures what?” cheat-sheets (short skims)
  • BrainFacts: “Measuring neural physiology”
  • BrainFacts: “Viewing the human brain through noninvasive tools”

Conditional Track B — Linear algebra intuition (refresher)

Complete this track if vectors/matrices feel rusty.
Target time: ~60 min total.

• 3Blue1Brown: “Vectors, what even are they?”
• 3Blue1Brown: “Linear combinations, span, and basis vectors”
• 3Blue1Brown: “Linear transformations and matrices”
• 3Blue1Brown: “Matrix multiplication as composition”

Conditional Track C — Deep learning intuition (refresher)

Complete this track if backprop/gradient descent are unfamiliar.
Target time: ~60 min total.

• 3Blue1Brown: “What is a neural network?”
• 3Blue1Brown: “Gradient descent”
• 3Blue1Brown: “Backpropagation”

Bootcamp Checkpoint (all students; due Sun 2/2, 2:00pm)

Reading: Crick (1989). The recent excitement about neural networks. Nature

Reading response due Sun 2/2, 2:00pm: 📝 Submit on Canvas

Optional background papers (only if you want extra framing)

• Cichy & Kaiser (2019). Deep Neural Networks as Scientific Models. Trends in Cognitive Sciences
• Doerig et al. (2023). The neuroconnectionist research programme. Nature Reviews Neuroscience

Architecture, learning objectives, and data shape representations, behavior, and generalization across modalities.

📑 Slides · 📺 Lecture capture

Reading (complete before class): Serre (2019). Deep Learning: The Good, the Bad, and the Ugly. Annual Review of Vision Science

Reading response due Tue 2/3, 2:00pm: 📝 Submit on Canvas

Student lightning talks on the three levers of deep learning (architecture, objectives, data).

📑 Lightning talk presentations · 📋 List of lightning talks (Google Sheet)

Reflection due Sun 2/8, 2:00pm: 📝 Submit on Canvas

Pretraining and fine-tuning/transfer; in-context learning and reasoning; what "emergence" claims mean and how to evaluate them critically.

📑 Slides

Reading (complete before class): Firestone (2020). Performance vs. competence in human–machine comparisons. PNAS

Reading response due Tue 2/10, 2:00pm: 📝 Submit on Canvas

Student lightning talks on scaling and emerging capabilities.

📋 List of lightning talks (Google Sheet)

No class.

Class on Feb 19 was cancelled (instructor sick).

Background paper (complete before or by Thu): Serre & Pavlick (2025). From Prediction to Understanding: Will AI Foundation Models Transform Brain Science? Neuron

Required viewing (watch both):

• Marcus, G. (2024). Keynote at AGI-24. Machine Learning Street Talk. Watch from ~5:00 to ~35:00 (Marcus's talk starts after a brief introduction).
• LeCun, Y. (2024). Objective-Driven AI: Towards AI Systems That Can Learn, Remember, Reason, and Plan. Ding Shum Lecture, Harvard CMSA. Watch the first ~36 minutes. Note: it gets fairly technical in places — if it starts feeling dry, you can stop around the 36-minute mark; you'll have what you need.

Paper response due Sun 2/22, 2:00pm: 📝 Submit on Canvas

Feature visualization, concept-based methods, sparse/dictionary approaches (incl. SAEs); what we can and can't reliably "name" in representations.

Guest lecturer: Dr Thomas Fel (Kempner Institute, Harvard) — thomasfel.fr

📑 Slides · 📺 Lecture capture

Reading (complete before class): Olah et al. (2018). The Building Blocks of Interpretability. Distill

Reading response due Tue 2/24, 2:00pm: 📝 Submit on Canvas

Student lightning talks on features, concepts, and sparse methods.

📑 Lightning talk presentations · 📋 List of lightning talks (Google Sheet)

Read: Thorpe (1989). Local vs. Distributed Coding. Intellectica

Reflection due Sun 2/29, 2:00pm: 📝 Submit on Canvas

Circuits, causal interventions, and standards of evidence for mechanistic claims.

Guest lecturer: Michael Lepori (CS, Brown) — lepori.xyz

📑 Slides · 📺 Lecture capture

Watch: Jensen Huang & Ilya Sutskever, Discovering the World Model Through LLM (YouTube)

Read: nostalgebraist (2020). Interpreting GPT: The Logit Lens. LessWrong

Reading response due Tue 3/3, 2:00pm: 📝 Submit on Canvas

Student lightning talks on circuits and causal interventions.

📑 Lightning talk presentations · 📋 List of lightning talks (Google Sheet)

Read: Anthropic. Tracing the Thoughts of a Large Language Model

Reflection due Sun 3/8, 2:00pm: 📝 Submit on Canvas

Predicting neural data across measurement modalities; encoding/decoding and representational similarity; what alignment can and cannot justify.

📑 Slides · 📺 Lecture capture

Reading (complete before class):

• Kriegeskorte & Douglas (2018). Cognitive computational neuroscience. Nature Neuroscience
• Doerig et al. (2023). The neuroconnectionist research programme. Nature Reviews Neuroscience

Reading response due Tue 3/10, 2:00pm: 📝 Submit on Canvas

Student lightning talks on model-to-brain mapping.

📑 Lightning talk presentations · 📋 List of lightning talks (Google Sheet)

Read: Linsley, Feng & Serre (2025). Better artificial intelligence does not mean better models of biology. Trends in Cognitive Sciences

Reflection due Sun 3/15, 2:00pm: 📝 Submit on Canvas

Treating models as participants in cognitive tasks; behavioral signatures beyond accuracy (generalization, planning, decision making, cognitive control); confounds and best practices.

📑 Slides · 📺 Lecture capture

Note: The beginning of the lecture is missing from this recording — the capture was not started until partway through class.

Reading (complete before class): Binz et al. (2025). A foundation model to predict and capture human cognition. Nature

Reading response due Tue 3/17, 2:00pm: 📝 Submit on Canvas

Student lightning talks on cognitive and behavioral evaluation.

📑 Lightning talk presentations · 📋 List of lightning talks (Google Sheet)

Read: Brady et al. (2025). Dual-process theory and decision-making in large language models. Nature Reviews Psychology

Reflection due Sun 3/22, 2:00pm: 📝 Submit on Canvas

No class

No class

Project jumpstart: walk through Tekin's demo notebook, pick a task from Psych-101, and begin running your replication pipeline. In-class working time.

📑 Slides · 📺 Lecture capture · 📋 Final project specification

Continue project work: complete runs and draft poster. TA sign-off required to access frontier models.

📋 Final project specification

Location: Kasper Multipurpose Room. Students present project findings in posters; structured peer feedback and synthesis discussion. Poster PDF due by 2:00 pm on Canvas.

📋 Final project specification

Location: Kasper Multipurpose Room. Students present project findings in posters; structured peer feedback and synthesis discussion. Peer feedback due end of day on Gradescope.

How the human brain represents and processes language—and what we learn by comparing brain and machine representations during language understanding.

Guest lecturer: Dr Greta Tuckute (Kempner Institute, Harvard) — tuckute.com

📑 Slides (PDF) · 📺 Lecture capture

Reading (complete before class): Tuckute, Kanwisher & Fedorenko (2024). Language in Brains, Minds, and Machines. Annual Review of Neuroscience

Reading response due Tue 4/14, 2:00pm: 📝 Submit on Canvas

Global workspace theory, consciousness, and deep learning.

Guest lecturer: Dr Rufin VanRullen (CNRS, France) — rufinv.github.io

📺 Lecture capture

Reading (complete before Thu lecture): VanRullen & Kanai (2021). Deep Learning and the Global Workspace Theory. Trends in Neurosciences

Reflection due Sun 4/19, 2:00pm: 📝 Submit on Canvas

Neuroscience-inspired recurrent models of vision: recurrent architectures motivated by cortical circuitry, how they go beyond feedforward DNNs, and links to behavioral and neural benchmarks (Serre lab).

Reading (complete before class): Read Kim et al. (2020) first, then Linsley et al. (2020). Prioritize main claims and Figures 1–2; skim implementation details.

• Kim, Linsley, Thakkar & Serre (2020). Disentangling neural mechanisms for perceptual grouping. ICLR (arXiv).
• Linsley, Kim, Ashok & Serre (2020). Recurrent neural circuits for contour detection (γ-net). ICLR (arXiv).

Reading response due Tue 4/21, 2:00pm: 📝 Submit on Canvas

Generative models, energy-based models, and cognitive science.

Guest lecturer: Dr Victor Boutin (CNRS, France) — victorboutin.github.io

No new reading for Thursday; come with questions for our guest.

Course-wrapping reflection due Sun 4/26, 2:00pm: 📝 Submit on Canvas