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.
3. New kind of methods
Recently, some methods integrate attention mechanism into the nerual network architectures to provide explanations.
Oppoturnities and Challenges
1. explanations of new AI architectures
Mechanistic interpretability can be an option to gain deep insights into generative models, like Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), Large Language Models (LLMs), etc.
2. improving XAI methods
Attribution methods have limitations. For example, attribution-based explanations are vulnerable to input perturbations, parameter custimization (selction of baselines), etc. One solution can be aggregating explanations of different XAI methods to gain a more accurate and robust explanation.
3. Concept-based methods
Concept-based learning methods are a popular class of approaches which could be used for both post-hoc and ante-hoc explanations. Many recent algorithms describe “prototypical concepts” or “prototypes”, like ProtoPNet, ProtoTree, ProtoPShare, Concept Bottleneck Models, concept Activation Vectors, Concept Embedding Models, Concept Atlases, etc. Neuro-symbolic learning and knowledge graphs are also popular methods fall within concept-based approaches.