Tejas Pradeep Shirodkar
PulseAugur coverage of Tejas Pradeep Shirodkar — every cluster mentioning Tejas Pradeep Shirodkar across labs, papers, and developer communities, ranked by signal.
-
New method measures singular structure in neural networks
Researchers have developed a new method to measure and classify the singular structure within trained neural networks without relying on descent or alignment. This technique, detailed in a new arXiv paper, allows for th…
-
New Dead-Direction Conditioners improve deep network optimization
Researchers have developed Dead-Direction Conditioners (DDC), a novel preconditioning method for deep neural networks designed to improve optimization stability and performance. DDC leverages gauge equivariance to keep …
-
New spectral method offers cheaper analysis of deep network complexity
Researchers have introduced Dead-Direction Signatures (DDS), a novel method for analyzing the complexity of deep neural networks. DDS offers a computationally cheaper alternative to existing techniques like the local le…
-
New diagnostic tool identifies 'dead directions' in LayerNorm transformers
Researchers have identified an algebraic method to detect 'dead directions' in LayerNorm transformers, which are parameter space directions where the Fisher information metric vanishes. This new diagnostic technique, de…