Gwen's Log

XAI research summary and trends

Tue, 04 Feb 2025 00:00:00 +0000

Methodologies

1. Attribution methods

Some methods contribute to the explanation of Deep Learning (DL) models by computing the importance/relevance of input features, like Local Interpretable Model-ggnostic Explanations (LIME), Shapley Additive Explanation (SHAP), etc. Some methods utilize network gradients, like saliency maos, Deconvolutional Neural Networks (DeConvNet), Layer-Wise Relevance Propagation (LRP), Pattern Attribution, and Randomized Input Sampling for Explanation (RISE), etc.

2. Ante-hoc explainable models

Traditional Machine Learning (ML) methods are often adopted for ante-hoc methods, especially techniques based on Decision Trees (DTs). Anothor recent development in XAI is rule-based approaches.

AI for weather and climate

Tue, 14 Jan 2025 00:00:00 +0000

1. ClimateLearn

[arxiv GitHub]

There are three tasks in ClimateLearn: Weather forecastiing, downscaling, and climate projection.

Datasets included: ERA5, CMIP6, PRISM

Models included: Baseline (Climatology, Persistence, Interpolation, Linear regression), and Deep Learning Models (ResNet, U-Net, Vision Transformer).

Mon, 01 Jan 0001 00:00:00 +0000

Jiawen WEI