Reading List 

This is a reading list intended to give a flavor of the types of problems in deep learning and AI that practicing (and erstwhile) physicists have worked on. 

Papers bracketed by ** are required reading, and appear in the schedule. 

Please note, that this is a selected literature review. A large amount of high quality science is not listed, but can be found if you  follow citations presented in the papers below.  

Part 1: Deep Learning Primer

Concepts in Deep Learning

Inspiration:

• Belkin, The Necessity of Machine Learning Theory in Mitigating AI Risk (2024)

A call to arms for physicsts to tackle just one of many existential risks to humanity

• Zdeborova, Understanding deep learning is also a job for physicists (2020)

Review papers

• **Mehta et al., A high-bias, low-variance introduction to Machine Learning for physicists (2019)**

• Bahri et al, Statistical Mechanics of Deep Learning (2020)

Philosophy

• Marr, Vision (1982), Ch. 1

Overparameterization and Double Descent

• Schaeffer et al., Double Descent Demystified (2023)

• Rocks and Mehta, Memorizing without overfitting: Bias, variance, and interpolation in over-parameterized models (2022) 

Feature vs Lazy Learning (how to scale deep nets)

• Yang, Simon, & Bernstein, A spectral conditioning for feature learning (2024)

• Karkada, The lazy (NTK) and rich (μP) regimes: a gentle tutorial (2024)

• Pehlevan and Bordelon, Lecture Notes on Infinite-Width Limits, Sec. 4

• Talk by Jamie Simon at KITP

Deep Learning Architectures

• Phuong and Hutter, Formal Algorithms for Transformers (2022)

For the most part, neural network architectures are described well in theory papers. This one is devoted exclusively to describing transformers, and is useful once you penetrate the pseudocode notation.

Statistical Mechanics of Learning

• Li and Sompolinsky, Statistical Mechanics of Deep Linear Neural Networks: The Backpropagating Kernel Renormalization (2021)

Language modeling

• Bengio et al, A neural probabilistic language model (2003)

Signal Propagation

Signal prop in feedforward MLPs + architectural variants

• Poole et al., Exponential expressivity in deep neural networks through transient chaos (2016)

• **Schoenholz et al, Deep Information Propagation (2017) **

• Doshi et al., Critical Initialization of Wide and Deep Neural Networks through Partial Jacobians: General Theory and Applications (2023)

RNNs

• Schmidhuber, Sepp Hochreiter's Fundamental Deep Learning Problem (1991) 

• Pascanu, Mikolov, Bengio, On the difficulty of training RNNs

• Molgedey et al., Suppressing Chaos in Neural Networks by Noise (1992)

• Krishnamurthy et al, Theory of Gating in RNNs (2022)

Transformers

• Dinan et al., Effective Theory of Transformers at Initialization (2023) 

• Cowsik et al., Geometric Dynamics of Signal Propagation Predict Trainability of Transformers (2024)  

Part 2: LLMs

Scaling Laws

Empirical

• **Kaplan et al., Scaling Laws for Neural Language Models (2020)**

• Henighan et al., Scaling Laws for Autoregressive Generative Modeling (2020)

• **Hoffmann et al., Training Compute-Optimal Large Language Models (Chinchilla) (2022) **

• OpenAI, GPT-4 technical report 

• Hernandez et al. (Anthropic), Scaling Laws and Interpretability of Learning from Repeated Data (2022)

Theory

• Bahri et al., Explaining Neural Scaling Laws (2024)

• Bordelon et al. A Dynamical Model of Neural Scaling Laws (2024)

• **Maloney et al., A Solvable Model of Neural Scaling Laws (2022)**

In-Context Learning

• **Brown et al. (OpenAI), Language Models are Few-shot learners (2020)** (GPT-3 paper)

• Anthropic, In-context learning and Induction Heads (2022)

• Lu et al., Asymptotic theory of in-context learning by linear attention (2024)

• Tong and Pehlevan, MLPs learn in-context (2024)

• Chan et al., Data Distributional Properties Drive Emergent In-Context Learning in Transformers (2022)

Emergence of Capabilities

• Anthropic, Predictability and Surprise

• Wei et al. Emergent abilities of LLMs (2022)

• Michaud et al., The quantization model of neural scaling (2024)

• Nam et al., An exactly solvable model for emergence and scaling laws (2024)

Hidden Capabilities

• Wei et al. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (2023)

• Suzgun et al., Challenging BIG-Bench with CoT (2022)

• Park et al. Emergence of Hidden Capabilities (2024)

• OpenAI o1

Grokking

• **Power et al., Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets (2022)**

• **Varma et al, Explaining Grokking through circuit efficiency (2023)**

• **Kumar et al., Grokking as the Transition from Lazy to Rich Training Dynamics (2024)**

• Huang et al., Unified View of Grokking, Double Descent and Emergent Abilities: A Comprehensive Study on Algorithm Task (2024)

Reinforcement Learning from Human Feedback (RLHF)

• Caspar, Davies et al., Open problems and fundamental limitations of RLHF

• Anthropic, Constitutional AI

• OpenAI, Training language models to follow instruction with human feedback

Mechanistic Interpretability

• Anthropic, Toy Models of Superposition

• Anthropic, Superposition, Memorization, and Double Descent

• Anthropic, Towards Monosemanticity