Research
I work on the science of deep learning. Most recently I have been thinking about synthetic data.
I flit between two largely disjoint communities who care about neural networks.
The first sees foundation models as oracles to automate the enterprise; this community thinks in terms of KV caches
and fast CUDA kernels, in terms of prefill and decoding and ensuring low latency user experiences for billions of
users around the world being exposed to the magic of next-token prediction for the first time. This one
might say a neural network is roughly a composed set of matrix multiplications done on tensor cores.
This community speaks to me because of the year I spent in the Bay Area working on software before college.
The second is comprised mostly of physicists and mathematicians who care about neural and cognitive systems.
This community sees neural networks as new Platonic objects. This community thinks in terms
of mean-field approximations and kernel methods, considering
neural networks as coupled units of computation, and high-dimensional loss (energy)
landscapes are the core object of inquiry.
This one would say neural networks are a type of parametric function class that is unexpectedly expressive, and
speaks to me because the phenomenology of using simple mathematical models to make non-obvious yet true predictions
about systems at scale is tantalizing.
I feel immensely lucky to have been mentored by people in both.
- Scaling Laws for Precision arXiv.
- Do Mice Grok? Unveiling Hidden Progress in Sensory Cortex During Overtraining arXiv.
- Asymptotic Dynamics for Delayed Feature Learning on a Toy Model HiLD at ICML 2024.
- No Free Prune: Information-Theoretic Barriers to Pruning at Initialization ICML 2024.
- Grokking as the Transition from Lazy to Rich Training Dynamics ICLR 2024.
- Human or Machine? Turing Tests for Vision and Language arXiv.
Tanishq Kumar*, Zachary Ankner*, Benjamin F. Spector, Blake Bordelon, Niklas Muennighoff, Mansheej Paul, Cengiz Pehlevan, Christopher Ré, Aditi Raghunathan.
Tanishq Kumar, Blake Bordelon, Cengiz Pehlevan, Venkatesh Murthy, Samuel J. Gershman.
Blake Bordelon, Tanishq Kumar, Samuel J. Gershman, and Cengiz Pehlevan.
Tanishq Kumar*, Kevin Luo*, Mark Sellke.
Tanishq Kumar, Blake Bordelon, Samuel J. Gershman*, Cengiz Pehlevan*.
Mengmi Zhang, ... Tanishq Kumar, ... Gabriel Kreiman.