ML Reading Group
Welcome
We are an informal machine learning reading group here at UC Berkeley. Our goal is to provide a low-stakes space for people to engage with and better understand modern ML research. We are structured around a topic of the month format. Each month, we pick a focused research area and read deeply into it, covering foundational papers, recent advances, and open questions. At every meeting, a subset of people volunteer to present a paper to the group. Those who don't present participate by asking questions and engaging in the group discussion.
We don't have formal prerequisites, but you'll get the most out of discussions if you have exposure to mathematical ML fundamentals, the kind covered in CS 189, CS 182, or similar. If you're newer to reading research papers, don't worry! This is a skill the group helps you build.
Overall, the group is pretty low-stakes and informal, and we are not affiliated with any orgs on campus or the department. If this sounds interesting to you, feel free to reach out using the link below!
Random Week: Final Meeting of the Year
Five talks. Five brilliant machine learners. Zero excuses to miss it.
- Neural Thickets — presented by guest speaker Prastik. Core idea: the neighborhood around pretrained weights contains task-specific experts recoverable with post-training RandOpt.
- Adjoint Matching — presented by Mark. Optimal control meets diffusion models.
- METRA — presented by Deepinder. A method aimed at making unsupervised RL scalable for broad downstream utility.
- Randomized Numerical Linear Algebra for ML — presented by Surya.
- The Muon Optimizer — presented by Dhruv. Known for strong practical performance, including NanoGPT speed-record interest.
Meeting time: Friday, 5-7 PM (reach out for the location). Come with questions, leave with knowledge (and snacks).
Recent Meetings
- Mar 20 · Reluplex: An Efficient SMT Solver for Verifying Deep Neural Networks
- Mar 20 · CoT May Be Highly Informative Despite "Unfaithfulness" (from METR)
- Feb 27 · Mastering Diverse Domains through World Models
- Feb 27 · Masked Contrastive Representation Learning for RL (M-CURL)
- Feb 13 · Contrastive Unsupervised Representations for RL (CURL)
- Feb 13 · Bootstrapped Representations in RL
- Feb 13 · Learning Invariant Representations for RL without Reconstruction
- Jan 26 · Native Parallel Reasoner (NPR)
Paper Selection Criteria
We prioritize papers that are:
- Foundational or highly influential within the topic area
- Quantitative, with clear empirical claims
- From well-known research groups or top venues (NeurIPS, ICML, ICLR, CVPR, etc.)
- Recent enough to reflect the current state of the field
Logistics
When: Fridays, 5–7 PM
Where: Contact Vijay for location by clicking the link below
About the Organizer
Hi everyone, my name's Vijay, and I'm a third-year computer science major.
I'm interested in natural language processing (NLP), machine learning systems, and AI in general.
Outside of school, I enjoy cooking and hiking.
vkethana@berkeley.edu