NObSP (Nonlinear Oblique Subspace Projections) is a framework that decomposes neural network predictions into explicit per‑feature contribution functions and an interaction residual, leveraging the linear final layer and oblique projections to avoid double counting when feature subspaces overlap. It connects to functional ANOVA and the Kolmogorov‑Arnold representation theorem, and introduces an efficient partial regression algorithm for out‑of‑sample evaluation. For convolutional networks, NObSP‑CAM generates class activation maps without backward passes after a single calibration, and experiments on tabular and vision datasets show faithfulness comparable to established attribution methods, with high function reproduction scores and improved class purity on TinyImageNet.
arXiv:2606. 01172v1 Announce Type: new Abstract: Modeling unknown latent functions from finite, irregularly sampled measurements is a recurring challenge across science and engineering.
arXiv:2310.16295v2 Announce Type: replace-cross Abstract: Neural network have achieved remarkable successes in many scientific fields. However, the interpretability of the neural network model is sti...
The paper introduces a control‑variable framework for deep neural networks to mitigate omitted variable bias, particularly shortcut learning where covariates like demographics influence predictions. It refits the final layer of a pre‑trained network using cross‑fitting with ridge penalisation, orthogonalises covariate effects, and marginalises predictions over covariate distributions to achieve unbiased, interpretable results. Experiments on simulated images and neuroimaging data show consistent estimation of true effects and performance close to models trained on unconfounded data.
arXiv:2609.06862v1 Announce Type: new Abstract: Superposition refers to neural networks representing more features than they have dimensions. It offers a possible explanation for polysemantic neurons...
arXiv:2609.40127v1 Announce Type: cross Abstract: Modern transformers pair impressive capabilities with substantial memory and compute demands. Low-rank weight factorization reduces both while keepin...
The paper investigates when supervised learning can recover the true latent variables that generate data, focusing on two causal settings: generative (labels cause latents) and discriminative (latents cause labels). It shows that if the representation space has at least as many dimensions as the latent space, optimal encoders are linear; otherwise, identifiability fails and encoders become nonlinear. Additional constraints—such as non‑negative activations in the generative case or sparsity and spark conditions in the discriminative case—enable recovery of individual latent variables, linking nonlinear ICA, compressed sensing, and neural network interpretability.
arXiv:2602. 00329v4 Announce Type: replace-cross Abstract: Reliable data attribution is essential for mitigating bias and reducing computational waste in modern machine learning, with the Shapley value serving as the theoretical gold standard.
arXiv:2606. 29951v1 Announce Type: new Abstract: Interpretable Mesomorphic Neural Networks (IMNs) offer a promising framework that combines the predictive power of deep neural networks with the interpretability of linear models.
arXiv:2607. 11990v1 Announce Type: cross Abstract: Feedforward network (FFN) blocks account for a large fraction of the parameters and computation in Transformer architectures, yet their internal structure remains difficult to interpret due to the additive superposition induced by the residual stream.
arXiv:2610.28555v1 Announce Type: cross Abstract: Statistical learning in high dimensions suffers from the curse of dimensionality. Fortunately, many functions of practical interest are compositional...
The paper revisits the Neural Thickets hypothesis, which claims that pretrained language models contain dense neighborhoods of task‑specific experts that can be accessed via random Gaussian weight perturbations. The authors reproduce the reported gains but demonstrate that similar improvements can be achieved by perturbing only a single decoder block or by keeping the model frozen and adding a random bias to its output logits. They propose the Perturb‑Select‑Vote (PSV) framework, which generates perturbed predictors, selects high‑performing ones on a small validation set, and aggregates their predictions, showing that output diversity, selection, and voting can explain the gains without relying on nearby weight‑space experts. PSV also improves vision tasks under domain shift, especially with logit perturbations, and offers a simple, training‑free alternative to full‑weight perturbation ensembles.
The paper introduces In-Table Prediction (ITB), a self‑supervised task where deep neural networks learn to predict any column in a table from the remaining columns. It proposes a novel neural layer to handle missing continuous values, generates synthetic datasets with controlled column relationships, and evaluates three architectures—MLP, ResNet, and Transformer—showing that attention‑based Transformers perform best when ample training data and large embeddings are used. The study is limited to synthetic, small‑column tables and is presented as an initial investigation rather than a comprehensive real‑world analysis.