Searcharxiv⌕ Search

arXiv subjects

Mitra Varun

Publications and source records attributed to Mitra Varun.

2 recordsLinked to original sources

Topological Dynamics Under Non-Stationary Supervision: Persistent Homology of Neural Network Weight Evolution

We investigate how the persistent topology of neuron- weight point clouds evolves under stable and non-stationary supervision. An identical multilayer perceptron is trained on Fashion-MNIST with correct labels, a fixed random permutation of the training-label vector, or a fresh permutation every epoch. Five paired seeds share initialisation and minibatch order, yielding 15 models trained for 100 epochs. Raw and row-normalised incoming weight clouds are analysed with H0 and H1 Vietoris Rips persistence at 22 checkpoints. Correct-label models reach 89.39% mean true-label test accuracy; fixed-random models fit 67.94% of their assigned training targets without generalising to the original task. Correct-label models have descriptively greater time-averaged normalised fc3 H1 persistence than either random-label condition, but the configured exact comparisons do not reach conventional significance after Holm correction. Early condition classification achieves 60.0% mean balanced accuracy using topology, 66.7% using conventional diagnostics, and 60.0% using both. Thus, no incremental predictive improvement is demonstrated in this setting. The archive contains all 3,300 configured topology cases. Interpretation is limited by five seeds, a single architecture and dataset, 282 capped H1 distance comparisons, and deferred bottleneck analysis.

cs.AI↗

ShiftGuard: Evaluating Confidence and Feature Distance for Risk-Controlled Image Classification

A classifier that is usually right still makes confident mistakes, and a deployed system needs some way to decide when to answer and when to abstain. Cheap post-hoc scores such as maximum softmax probability (MSP) and predictive entropy are the standard answer. Distances in feature space are a popular alternative, and fusing several signals into one score is a natural next step. We ask whether adding a feature distance signal to MSP and entropy improves error detection, and whether it increases the share of predictions that can be accepted at a certified error rate when only a small labelled calibration set is available. We train a CIFAR-adapted ResNet-18 from scratch on CIFAR-10 (92.92% test accuracy, one seed) and compare five reliability scores: MSP, entropy, distance to the predicted-class centroid, a logistic fusion of MSP and entropy, and a logistic fusion of all three signals. Acceptance thresholds are chosen on a held out calibration set with Clopper Pearson upper bounds, Bonferroni corrected over a prespecified family of 315 rules at 95% simultaneous confidence. Confidence based scores were best. Error AUROC was 0.9230 for MSP, 0.9233 for entropy and 0.9235 for MSP plus entropy, against 0.8814 for the three signal fusion and 0.8224 for distance alone. At a 5% risk target with 1,000 calibration images, entropy accepted 84.82% of test images with 2.05% observed error, while three signal fusion accepted 80.61% with 2.21%. With 250 calibration images, neither distance nor the three-signal fusion could certify the 5% target at all. The results come from one seed, one dataset and one architecture, with no corruption experiments, so they say nothing about distribution shift detection. Within these limits, the distance signal did not earn its added complexity.

cs.CV↗