Parameterized synthetic text generation with simplestories
A dataset full of simple yet diverse stories; the MNIST for language
A taxonomy of interpretation and explanation methods for capsule network architectures
This paper presents a comprehensive taxonomy of interpretation and explanation methods developed for capsule network (capsnet) architectures, analyzing their mechanisms, applicability, and performance across diverse problem domains.
Finding manifolds with bilinear autoencoders
Decomposing activations into sparse polynomials and using their geometry
Smooth infomax - towards easier post-hoc interpretability
Sim makes post-hoc interpretability tools more effective through latent space constraints
Improving neural network accuracy by concurrently training with a twin network
We show the applicability of twin network augmentation on convolutional neural networks
Label-efficient learning for radio frequency fingerprint identification
We introduce a label-efficient approach for radio frequency fingerprint identification, achieving competitive accuracy with up to 10x fewer labels.
Towards the characterization of representations learned via capsule-based network architectures
This paper provides a systematic and principled study on the interpretability of capsule network (capsnet) representations, aiming to characterize the nature and structure of the learned features across diverse architectures and datasets
Analyzing the explanation and interpretation potential of matrix capsule networks
This study investigates the internal mechanisms of matrix capsule networks with the EM routing algorithm
Compositionality unlocks deep interpretable models
Introducing a global svd-like algorithm for multi-linear models.
Bilinear MLPs enable weight-based mechanistic interpretability
Using bilinear MLPs to reverse-engineer shallow MNIST and tiny stories models from their weights.
Twin network augmentation
A novel training strategy for improved spiking neural networks and efficient weight quantization.
Tokenized SAEs: disentangling SAE reconstructions
We use a per-token bias in SAEs to separate token reconstructions from interesting features.
Weight-based decomposition: a case for bilinear MLPs
Introducing bilinear MLPs as a new approach to weight-based interpretability.
A contrastive learning method for multi-label predictors on hyperspectral images
Self-supervised contrastive learning for multi-label hyperspectral image classification
Training methods of multi-label prediction classifiers for hyperspectral remote sensing images
A deep learning model for hyperspectral remote sensing, shifting from traditional single-label, pixel-level classification to multi-label, patch-level analysis
The trifecta: three techniques for deeper forward-forward networks
Three techniques to significantly improve the forward-forward algorithm. We achieve 84% on CIFAR-10.
Gated recurrent unit with multilingual universal sentence encoder for arabic aspect-based sentiment analysis
This study presents a deep learning model for arabic aspect-based sentiment analysis (ABSA) using gated recurrent units (gru) combined with features from the multilingual universal sentence encoder (muse)