Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
Temporal Salience
Technique that identifies the most influential moments or time intervals in a time series by perturbing input data and measuring the impact on the model's output.
Temporal Guided Backpropagation
Interpretability method that adapts guided backpropagation to recurrent neural networks to visualize the temporal features that most activate neurons.
LIME for Time Series
Adaptation of the LIME (Local Interpretable Model-agnostic Explanations) algorithm that generates local explanations by creating perturbed segments of the time series to form a simple interpretable model.
Temporal SHAP
Extension of SHapley Additive exPlanations values to sequential data, attributing a contribution to each time step or each feature at each instant to explain the overall prediction.
Temporal Interval Masking
Interpretability approach that masks or replaces entire segments of the time series to assess their collective importance in the model's decision, unlike salience which focuses on individual points.
Temporal Evolution Rules
Method that extracts logical rules describing how states or patterns evolve over time to lead to a specific prediction, making the model's sequential reasoning explicit.
Latent Space Trajectory Analysis
Technique that visualizes and interprets the path taken by a data sequence in the latent space of a model (such as an autoencoder) to understand its dynamics and classification.
Temporal Relevance via Wavelet Decomposition
Method that decomposes the time series into different frequencies and time scales via wavelets, then evaluates the relevance of each component for the model's prediction.
Temporal Counterfactual Explanations
Generation of minimally modified time series that change the model's prediction, enabling understanding of critical temporal conditions that would have led to a different outcome.
Dynamic Heat Map
Interpretability visualization that displays the importance of features (or pixels in a video) in an evolving manner over time, showing how the model's focus changes.
Interpretability by Aggregation of Temporal Features
Approach that explains predictions based on aggregated temporal features (moving average, variance, etc.) rather than raw data points, offering a more macroscopic view.
Time Step Influence Decomposition
Method that isolates the contribution of each individual time step to the final prediction, often using denoising techniques or analyzing gradients through recurrent steps.
Long-Term Dependency Importance Analysis
Set of techniques aimed at quantifying and visualizing how distant events in the past influence the current prediction, a key challenge for models like LSTM or Transformers.
Temporal Causal Explanations
Methodology that goes beyond correlation to identify cause-effect relationships in sequential data that are exploited by the model, using causal models like temporal causal acyclic graphs.