no code implementations • 2 Feb 2024 • Tanishq Kumar, Kevin Luo, Mark Sellke
We put forward a theoretical explanation for this, based on the model's effective parameter count, $p_\text{eff}$, given by the sum of the number of non-zero weights in the final network and the mutual information between the sparsity mask and the data.
no code implementations • 9 Oct 2023 • Tanishq Kumar, Blake Bordelon, Samuel J. Gershman, Cengiz Pehlevan
We identify sufficient statistics for the test loss of such a network, and tracking these over training reveals that grokking arises in this setting when the network first attempts to fit a kernel regression solution with its initial features, followed by late-time feature learning where a generalizing solution is identified after train loss is already low.
no code implementations • 23 Nov 2022 • Mengmi Zhang, Giorgia Dellaferrera, Ankur Sikarwar, Marcelo Armendariz, Noga Mudrik, Prachi Agrawal, Spandan Madan, Andrei Barbu, Haochen Yang, Tanishq Kumar, Meghna Sadwani, Stella Dellaferrera, Michele Pizzochero, Hanspeter Pfister, Gabriel Kreiman
To address this question, we turn to the Turing test and systematically benchmark current AIs in their abilities to imitate humans.