Please enable JavaScript.

Coggle requires JavaScript to display documents.

Проблема: Variable input saliency - Coggle Diagram

- - - - Precision (max=1): полученные фичи должны содержать meaningful signals (сколько из найденных фичей на самом деле важны)
        
        Are all features identified as salient informative?
      - Recall (max=1): количество фичей, которые содержат important signal (доля найденных их всех -- находятся не все)
        
        Was the saliency method able to identify all informative features?
      - Some neural architectures tend to use more feature information when making a prediction (i.e., have more recall in terms of importance); this may be the desired behavior in many time series applications where importance changes over time, and the goal of using an interpretability measure is to detect all relevant features across time.
        
        Нельзя сравнивать методы просто по изменению prediction после masking
    - - Perturbation based
        
        Feature occlusion
        
        Differnce in output when REPLACING CONTIGUOUS REGION WITH BASELINE
        
        For time series we considered continuous regions as features with in same time step or multiple time steps grouped together.
        
        Feature Ablation
        
        computes attribution as the difference in output after replacing each feature with a baseline
        
        Feature Permutation
        
        Random feature permute within a batch
      - Gradient based
        
        Integrated gradients
        
        Average gradient while input changes from baseline
        
        Smoothgrad
        
        Gradient computed N times with noise each time
        
        DeepLIFT
        
        Neuron activation?
        
        Gradient SHAP
      - TSR on top of any saliency method
      - Baseline: random assignment
      - Модели
        
        LSTM
        
        LSTM with input-cell attention
        
        TCN
        
        Трансформеры