Gregory Valiant

• >main
• >papers
• >teaching
• >code
• >contact

Code

What Can Transformers Learn In-Context? A Case Study of Simple Function Classes. (With Shivam Garg, Dimitris Tsipras, and Percy Liang, NeurIPS, 2022):

[Code and Models] [paper]

Implicit regularization for deep neural networks driven by an Ornstein-Uhlenbeck like process. (With Guy Blanc, Neha Gupta, and Paul Valiant, COLT, 2020):

[ Code for generating figure 2] [ paper]

Spectrum Estimation (w. Weihao Kong, Annals of Statistics 2017):

Code for accurately recovering the eigenvalues of the covariance of a high-dimensional distribution, given access to a data matrix consisting of samples drawn from the distribution. Even in the regime where the dimensionality of the data exceeds the sample size, accurate recovery is possible. Code here, paper here.

Automatic Inequality Prover (w. Paul Valiant, FOCS 2014):

code here, See the paper here.

Estimating the Unseen (w. Paul Valiant):

Code for recovering an accurate approximation of the "histogram" of a distribution, given access to independent samples. This can be used to estimate statistical properties, such as support size, entropy, etc. Code here, paper here. For large-scale instances, also see the more efficient (and slightly more accurate) code here.