Yorumlanabilirlik

Inside The Model

1. Probing RepresentationsTraining simple classifiers on hidden states to test what information a layer encodes.1–2 sa
2. Mechanistic Interpretability: CircuitsReverse-engineering the algorithms inside networks: features, circuits, and causal interventions such as activation patching.1–2 sa
3. Superposition and Sparse Dictionary FeaturesNetworks pack more features than dimensions; sparse autoencoders try to recover interpretable features.1–2 sa

Explaining Predictions

1. Intrinsically Interpretable ModelsModels whose structure is the explanation: sparse linear models, shallow trees, and rule lists.1–2 sa
2. Feature Attribution: Permutation Importance and SHAPModel-agnostic explanations: how much predictions depend on each feature, globally and for one example.1–2 sa
3. Gradient-Based Saliency and Integrated GradientsAttributing a neural network's output to input pixels or tokens via gradients along a path from a baseline.1–2 sa