The field of artificial intelligence is moving towards developing more explainable and interpretable models. Recent research has focused on improving the transparency of neural networks, with a particular emphasis on techniques such as concept probing, sparse information disentanglement, and counterfactual explanations. These methods aim to provide insights into the decision-making processes of complex models, making them more trustworthy and reliable. Noteworthy papers in this area include SIDE, which introduces a novel method for improving the interpretability of prototypical parts-based neural networks, and Compositional Function Networks, which proposes a framework for building inherently interpretable models by composing elementary mathematical functions. Overall, the field is shifting towards a more nuanced understanding of neural network behavior, with a focus on developing techniques that can provide clear and concise explanations for model decisions.
Advances in Explainable AI and Neural Network Interpretability
Sources
Concept Probing: Where to Find Human-Defined Concepts (Extended Version)
Higher-order Kripke models for intuitionistic and non-classical modal logics
SIDE: Sparse Information Disentanglement for Explainable Artificial Intelligence
Counterfactual Explanations in Medical Imaging: Exploring SPN-Guided Latent Space Manipulation
Neural Estimation of the Information Bottleneck Based on a Mapping Approach
A Tensor-Based Compiler and a Runtime for Neuron-Level DNN Certifier Specifications
Technical Indicator Networks (TINs): An Interpretable Neural Architecture Modernizing Classic al Technical Analysis for Adaptive Algorithmic Trading
On the Limits of Hierarchically Embedded Logic in Classical Neural Networks
Compositional Function Networks: A High-Performance Alternative to Deep Neural Networks with Built-in Interpretability
On Explaining Visual Captioning with Hybrid Markov Logic Networks
What Does it Mean for a Neural Network to Learn a "World Model"?
Unlocking Interpretability for RF Sensing: A Complex-Valued White-Box Transformer
Explaining Deep Network Classification of Matrices: A Case Study on Monotonicity
Deep learning of geometrical cell division rules
Teaching the Teacher: Improving Neural Network Distillability for Symbolic Regression via Jacobian Regularization
Tapping into the Black Box: Uncovering Aligned Representations in Pretrained Neural Networks
An Information Bottleneck Asset Pricing Model
Causal Identification of Sufficient, Contrastive and Complete Feature Sets in Image Classification
I Am Big, You Are Little; I Am Right, You Are Wrong