סמינר: Graduate Seminar
Network Probing and Diagnostics through Null Space Analysis
Date:
September,14,2025
Start Time:
11:00 - 12:00
Location:
506, Zisapel Building
Add to:
Lecturer:
Harel Yadid
Research Areas:
| In modern machine learning, neural architectures, composed of learned weights and nonlinearities, achieve strong performance on classification, decoding, detection, and other vision tasks. However, the path by which these systems reach a decision often remains opaque. Standard methods in explainable AI (XAI) usually focus on features which positively affect the network’s prediction. Yet, classifiers contain many directions that leave the prediction unchanged. Such directions appear in the null space of latent features within the network’s layers. Identifying and visualizing these invariants is essential to understand what a model attends to and what it ignores. We present a method that probes the null space to expose invariant directions, visualize them, and use the resulting evidence as a diagnostic lens on neural networks. This procedure clarifies why certain changes in the input are immaterial to the classifier, while seemingly modest perturbations can be decisive. In order to achieve human-interpretable diagnostics, we translate features to a semantic multimodal space, such as CLIP, to bridge visual features and language guidance. The study gives a practical tool for invariant interpretability, reveals actionable directions, and enables inspection across layers of encapsulated information. |
| student Under the supervision of Prof. Guy Gilboa. |