Subspace Crystallization: Grokking in Modular Arithmetic Networks
submitted to NeurIPS'26
ICML Mech Interp Wkshp'26
Shlok Mehendale, Paras Chopra · Lossfunk
A causally active subspace crystallises inside the weights while the rest stays inert and free. Compression follows generalisation, never causes it.
submitted to ECCV'26
Sravan Danda, Aditya Challa, Shlok Mehendale, Snehanshu Saha · APPCAIR
most test-time adaptation methods are tied to specific architectures. this one matches geometric quantiles of classifier features to correct for distribution shift, works on transformers and convnets alike, no labelled corrupted data required.
submitted to TMLR
Shlok Mehendale, Aditya Challa, Rahul Yedida, Sravan Danda, Santonu Sarkar, Snehanshu Saha · APPCAIR
the right loss function for anomaly detection is the vanilla loss reweighted by the Radon-Nikodým derivative, derived from the normal and anomalous distributions. PAC learnability established, evaluated across 96 datasets.
submitted to IEEE TAI
Shanay Mehta*, Shlok Mehendale*, Nicole Fernandes*, Jyotirmoy Sarkar, Santonu Sarkar, Snehanshu Saha · cited by IBM Research
30+ algorithms across 104 datasets: tree-based methods match or beat deep learning on small univariate data while being orders of magnitude faster.
Stereotypy-Driven Exploration: Linking Habit and Flexibility in dlPFC–Striatum Circuits
In progress
Shlok Mehendale, Upendra Vishwanath, Prasanna Venkhatesh
We propose a repetition-sensitive reinforcement learning framework where the striatum encodes reinforcement history to drive habitual exploitation and the dlPFC monitors stereotypy to trigger exploratory shifts, positioning repetition itself rather than novelty or uncertainty as the key biological signal for fronto-striatal flexibility.
Weighted Attention Bottleneck Transformer for Multimodal Fusion
UofT'25
Shlok Mehendale · Won ProjectX · ETH Zurich, Cornell, CMU, Edinburgh
cross-modal reconstruction via a hierarchical bottleneck; adaptive fusion without full cross-attention, building on Google Research's MBT.
submitted to Comp. Intelligence
Shlok Mehendale, JK Sahoo · BITS Pilani
treats haze as a modelling problem, not a preprocessing step — CNN–PinGTSVM achieves 98.20% vs. 88.63% for dehazing-first baselines.