InversionView: A Method for Understanding Information Encoded in Neural Network Activations
InversionView is a novel method for deciphering the information encoded within neural network activations by sampling potential inputs from a trained decoder model, offering insights into the inner workings of these complex systems.
Improving the Interpretability of Neural Networks Using Sparse Autoencoders with Mutual Feature Regularization
Sparse autoencoders (SAEs) can be enhanced to learn more interpretable features by encouraging multiple SAEs trained in parallel to learn similar features, a technique called Mutual Feature Regularization (MFR), leading to improved reconstruction loss and better preservation of input feature information.
Interpreting Neural Network Linear Layers as Approximations of Mahalanobis Distance Using Absolute Value Activations
This paper proposes a novel theoretical framework for interpreting neural networks by establishing a mathematical connection between linear layers with Absolute Value (Abs) activations and the Mahalanobis distance, a statistical measure accounting for data covariance.
Identifying Functionally Distinct Sub-networks within Neural Networks Using Gromov-Wasserstein Distance
This paper introduces a novel, automated, and task-agnostic method for identifying functionally distinct sub-networks within neural networks by leveraging the Gromov-Wasserstein (GW) distance to measure functional similarity between intermediate layer representations.
Unlearning-Based Neural Interpretations for Improved Attribution in Image Classification
Unlearning-based Neural Interpretations (UNI) offers a novel approach to improve the faithfulness, stability, and robustness of gradient-based attribution methods in deep neural networks by generating debiased and adaptive baselines through targeted unlearning.
The Computational Complexity of Discovering Interpretable Circuits in Neural Networks
Discovering functionally meaningful circuits within neural networks, a key aspect of inner interpretability, is computationally challenging, often even intractable, demanding a nuanced understanding of the complexity landscape and exploration of viable algorithmic options.
© 2024 by Linnk AI