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.