The Holistic Interpretability

Analyzed the CNNs trained on MNIST and FashionMNIST using neuron ablation, causal tracing and activation interventions to understand the role of individual neurons and their interactions in model predictions.

Research Project Details

Add detailed information about your research project here.

Abstract

Brief summary of the research.

Methodology

Describe your approach.

Results

Across all experiments, LLMs produced plausible yet inaccurate estimates, showing that they rely on heuristics rather than actual computation.