About the course
A seminar on NeuroAI, the joint study of artificial and biological intelligence, for advanced undergraduates and graduate students. The course covers how modern deep networks and foundation models are trained, what interpretability methods reveal about them, and how well they match brains and behavior.
Each topic has required reading for everyone. Students then present recent papers on the topic in short talks.
Topics
- Foundations
- The three levers of deep learning
- Scaling and emerging capabilities
- Prediction vs. understanding
- Representation-level interpretability
- Mechanistic interpretability
- Neural alignment
- Behavioral and cognitive alignment
- Language and consciousness
- Feedback in biological and artificial vision
Foundations
Why neural networks became models of the brain, and what that framing assumes.
Required reading
- The recent excitement about neural networksCrick · Nature 1989
- Deep neural networks as scientific modelsCichy & Kaiser · Trends in Cognitive Sciences 2019
The three levers of deep learning
How architecture, learning objective and training data shape what a network learns and how it generalizes.
Required reading
- Deep learning: the good, the bad, and the uglySerre · Annual Review of Vision Science 2019
Student presentations
Architectures
- ImageNet classification with deep convolutional neural networksKrizhevsky, Sutskever & Hinton · NeurIPS 2012
- Deep residual learning for image recognitionHe et al. · CVPR 2016
- Attention is all you needVaswani et al. · NeurIPS 2017
- BERT: Pre-training of deep bidirectional transformers for language understandingDevlin et al. · NAACL 2019
- Language models are unsupervised multitask learnersRadford et al. · OpenAI 2019
- An image is worth 16x16 words: Transformers for image recognition at scaleDosovitskiy et al. · ICLR 2021
Datasets and platforms
- ImageNet large scale visual recognition challengeRussakovsky et al. · International Journal of Computer Vision 2015
- Habitat: A platform for embodied AI researchSavva et al. · ICCV 2019
- Exploring the limits of transfer learning with a unified text-to-text transformerRaffel et al. · JMLR 2020
- The Pile: An 800GB dataset of diverse text for language modelingGao et al. · arXiv 2021
- LAION-5B: An open large-scale dataset for training next generation image-text modelsSchuhmann et al. · NeurIPS 2022
Self-supervised learning
- A simple framework for contrastive learning of visual representationsChen et al. · ICML 2020
- Learning transferable visual models from natural language supervisionRadford et al. · ICML 2021
- Emerging properties in self-supervised vision transformersCaron et al. · ICCV 2021
- Masked autoencoders are scalable vision learnersHe et al. · CVPR 2022
Scaling and emerging capabilities
Pretraining, fine-tuning, in-context learning and reasoning, and how to evaluate claims of emergence.
Required reading
- Performance vs. competence in human–machine comparisonsFirestone · PNAS 2020
Student presentations
Pretraining and transfer
- Language models are few-shot learnersBrown et al. · NeurIPS 2020
- Training language models to follow instructions with human feedbackOuyang et al. · NeurIPS 2022
In-context learning and reasoning
- Rethinking the role of demonstrations: What makes in-context learning work?Min et al. · EMNLP 2022
- Parallel trade-offs in human cognition and neural networks: The dynamic interplay between in-context and in-weight learningRussin, Pavlick & Frank · PNAS 2025
- Chain-of-thought prompting elicits reasoning in large language modelsWei et al. · NeurIPS 2022
- DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learningGuo et al. · Nature 2025
Vision and language
- Do vision and language models share concepts? A vector space alignment studyLi et al. · TACL 2024
- Visual instruction tuningLiu et al. · NeurIPS 2023
Critical perspectives
- Dissociating language and thought in large language modelsMahowald et al. · Trends in Cognitive Sciences 2024
- On the dangers of stochastic parrots: Can language models be too big?Bender et al. · FAccT 2021
- The debate over understanding in AI's large language modelsMitchell & Krakauer · PNAS 2023
Prediction vs. understanding
Whether models that predict behavior and brain activity also explain them.
Required reading
- From prediction to understanding: Will AI foundation models transform brain science?Serre & Pavlick · Neuron 2025
Representation-level interpretability
Feature visualization, concept-based methods and sparse dictionary learning. Guest lecture by Thomas Fel (Kempner Institute, Harvard).
Required reading
- The building blocks of interpretabilityOlah et al. · Distill 2018
- Local vs. distributed codingThorpe · Intellectica 1989
Student presentations
Feature visualization and circuits
- Feature visualizationOlah, Mordvintsev & Schubert · Distill 2017
- Zoom in: An introduction to circuitsOlah et al. · Distill 2020
- Multimodal neurons in artificial neural networksGoh et al. · Distill 2021
Attribution and concepts
- Grad-CAM: Visual explanations from deep networks via gradient-based localizationSelvaraju et al. · ICCV 2017
- Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (TCAV)Kim et al. · ICML 2018
- CRAFT: Concept recursive activation factorization for explainabilityFel et al. · CVPR 2023
- Concept bottleneck modelsKoh et al. · ICML 2020
Sparse representations
- Sparse autoencoders find highly interpretable features in language modelsCunningham et al. · ICLR 2024
- Scaling monosemanticity: Extracting interpretable features from Claude 3 SonnetTempleton et al. · Anthropic 2024
- Into the rabbit hull: From task-relevant concepts in DINO to Minkowski geometryFel et al. · arXiv 2025
Mechanistic interpretability
Probing, circuits and causal interventions. Guest lecture by Michael Lepori (Brown).
Required reading
- Interpreting GPT: The logit lensnostalgebraist · LessWrong 2020
- Tracing the thoughts of a large language modelAnthropic · Anthropic 2025
Student presentations
Probing
- A structural probe for finding syntax in word representationsHewitt & Manning · NAACL 2019
- BERT rediscovers the classical NLP pipelineTenney, Das & Pavlick · ACL 2019
- Racing thoughts: Explaining contextualization errors in large language modelsLepori, Mozer & Ghandeharioun · NAACL 2025
Causal methods and concept erasure
- Causal abstractions of neural networksGeiger et al. · NeurIPS 2021
- Interpretability in the wild: A circuit for indirect object identification in GPT-2 smallWang et al. · ICLR 2023
- Null it out: Guarding protected attributes by iterative nullspace projectionRavfogel et al. · ACL 2020
World models and relations
- Emergent world representations: Exploring a sequence model trained on a synthetic taskLi et al. · ICLR 2023
- Can language models encode perceptual structure without grounding? A case study in colorAbdou et al. · CoNLL 2021
- Beyond the doors of perception: Vision transformers represent relations between objectsLepori et al. · NeurIPS 2024
- On emergent social world models: Evidence for functional integration of theory of mind and pragmatic reasoning in language modelsTsvilodub et al. · arXiv 2026
Neural alignment
Comparing model representations with neural recordings, and what a good fit does and does not show.
Required reading
- Cognitive computational neuroscienceKriegeskorte & Douglas · Nature Neuroscience 2018
- The neuroconnectionist research programmeDoerig et al. · Nature Reviews Neuroscience 2023
- Better artificial intelligence does not mean better models of biologyLinsley, Feng & Serre · Trends in Cognitive Sciences 2026
Student presentations
Vision
- Performance-optimized hierarchical models predict neural responses in higher visual cortexYamins et al. · PNAS 2014
- Brain-Score: Which artificial neural network for object recognition is most brain-like?Schrimpf et al. · bioRxiv 2018
- Diverse deep neural networks all predict human inferior temporal cortex well, after training and fittingStorrs et al. · Journal of Cognitive Neuroscience 2021
Language and speech
- The neural architecture of language: Integrative modeling converges on predictive processingSchrimpf et al. · PNAS 2021
- Evidence of a predictive coding hierarchy in the human brain listening to speechCaucheteux, Gramfort & King · Nature Human Behaviour 2023
- Shared computational principles for language processing in humans and deep language modelsGoldstein et al. · Nature Neuroscience 2022
- Driving and suppressing the human language network using large language modelsTuckute et al. · Nature Human Behaviour 2024
- Dissecting neural computations in the human auditory pathway using deep neural networks for speechLi et al. · Nature Neuroscience 2023
- Decoding speech perception from non-invasive brain recordingsDéfossez et al. · Nature Machine Intelligence 2023
Foundation models for neuroscience
- A unified, scalable framework for neural population decodingAzabou et al. · NeurIPS 2024
- Foundation model of neural activity predicts response to new stimulus typesWang et al. · Nature 2025
Behavioral and cognitive alignment
Testing models as participants in cognitive experiments.
Required reading
- A foundation model to predict and capture human cognitionBinz et al. · Nature 2025
- Dual-process theory and decision-making in large language modelsBrady et al. · Nature Reviews Psychology 2025
Student presentations
Evaluation frameworks
- Large language models pass the Turing testJones & Bergen · arXiv 2025
- Triangulating LLM progress through benchmarks, games, and cognitive testsMomentè et al. · Findings of EMNLP 2025
- Levels of analysis for large language modelsKu et al. · arXiv 2025
Reasoning, analogy and social cognition
- Emergent analogical reasoning in large language modelsWebb, Holyoak & Lu · Nature Human Behaviour 2023
- Using counterfactual tasks to evaluate the generality of analogical reasoning in large language modelsLewis & Mitchell · CogSci 2024
- Position: Theory of mind benchmarks are broken for large language modelsRiemer et al. · ICML 2025
Decisions and biases
- Cognitive bias in decision-making with LLMsEchterhoff et al. · Findings of EMNLP 2024
- Embers of autoregression show how large language models are shaped by the problem they are trained to solveMcCoy et al. · PNAS 2024
Perception
- Large language models predict human sensory judgments across six modalitiesMarjieh et al. · Scientific Reports 2024
- Object-centric binding in contrastive language-image pretrainingAssouel et al. · arXiv 2025
Responses to Centaur
- "Captured" by Centaur: Opaque predictions or process insights?Kieval & Buckner · Journal of Experimental Psychology: Animal Learning and Cognition 2026
- Not yet AlphaFold for the mind: Evaluating Centaur as a synthetic participantNamazova et al. · arXiv 2025
Language and consciousness
Guest lectures by Greta Tuckute (Kempner Institute, Harvard) and Rufin VanRullen (CNRS, Toulouse).
Required reading
- Language in brains, minds, and machinesTuckute, Kanwisher & Fedorenko · Annual Review of Neuroscience 2024
- Deep learning and the global workspace theoryVanRullen & Kanai · Trends in Neurosciences 2021
Feedback in biological and artificial vision
Recurrent models of vision motivated by cortical circuits. Guest lecture by Victor Boutin (CNRS, Toulouse) on generative models in cognitive science.
Required reading
- Disentangling neural mechanisms for perceptual groupingKim, Linsley, Thakkar & Serre · ICLR 2020
- Recurrent neural circuits for contour detectionLinsley, Kim, Ashok & Serre · ICLR 2020