CORTEXA
← Browse

Muhammad Hamza

2 papers indexed

arxivcs.LGmath.OCstat.ML2026-06-29

Curvature-Weighted Gradient Diversity: A Noise Measure for Geometry-Adaptive SGD Schedules

Muhammad Hamza, Ayush Goel

The standard convergence analysis of mini-batch stochastic gradient descent (SGD) models gradient noise using a single variance term that treats all parameter directions equally, ignoring the fact that noise in high-curvature directions has less impact because learning rates are…

View free PDFSource page
arxivcs.NEcs.LG2026-06-26

Criticality-Constrained Iterative Pruning for Energy-Efficient Spiking Neural Networks via Combined Importance Scoring

Muhammad Hamza

Deploying spiking neural networks (SNNs) on neuromorphic hardware demands aggressive synaptic pruning while preserving temporal computation integrity. Existing strategies either neglect neuronal criticality or rely on convex relaxations of the inherently combinatorial pruning pro…

View free PDFSource page