Research on neural architectures, representations, and learning dynamics. Explore evidence about how network design affects capabilities and efficiency.
22 papers · Latest edition September 29, 2026
Three of the latest briefs in this collection. Read the evidence and the original papers alongside them.
A new query-efficient technique makes it possible to steal the parameters of small black-box neural networks using only their predicted labels.
Stembed lets music producers search for individual instrument sounds hidden inside a full song by representing the mixture as multiple searchable stem-like embeddings.
This paper argues that claims about linear representations only become meaningful once we specify which transformations leave a representation essentially unchanged.
Newest editions first.
Akira Ito, Takayuki Miura, Yosuke Todo
A new query-efficient technique makes it possible to steal the parameters of small black-box neural networks using only their predicted labels.
David Braun, Junyi Fan, Pranay Manocha, Donald S. Williamson, Adam Finkelstein
Stembed lets music producers search for individual instrument sounds hidden inside a full song by representing the mixture as multiple searchable stem-like embeddings.
Louie Hong Yao, Yuhao Li, Shengchao Liu
This paper argues that claims about linear representations only become meaningful once we specify which transformations leave a representation essentially unchanged.
Namwoo Kim, Hyungryul Baik, Yoonjin Yoon
HypLTSF maps multi-scale time-series patterns into hyperbolic space so that fine-to-coarse temporal hierarchies become explicit and useful for long-term forecasting.
Hanwen Jiang
The paper helps recurrent networks learn reliably far beyond the sequence lengths they saw during training by stabilizing how future errors credit earlier states.
Shivang Rawat, Mirko Morello, Flaviano Morone, David J. Heeger
This work shows that anticipating bottom-up signals can make deep continuous-time recurrent networks learn better with fewer parameters.
Jaeyong Bae, Hawoong Jeong
The study shows that more complex datasets reshape where low-loss parameter neighborhoods contract around trained neural networks.
Chun-Hua Lin, Samuel Yen-Chi Chen, Yu-Chao Hsu, Kuo-Chung Peng, Jiun-Cheng Jiang, Chi-Sheng Chen, Tai-Yue Li, Nan-Yow Chen, En-Jui Kuo, Hsi-Sheng Goan
A compact quantum-inspired neural network improves federated arrhythmia detection while reducing communication costs across hospital and wearable-device clients.
Jiaming Qiu, Yingye Zheng, Ying-Qi Zhao
A recurrent neural method learns when to classify patients early while balancing accuracy, sensitivity, specificity, and monitoring costs with explicit performance guarantees.
Liu Ziyin, Yizhou Xu, Tomaso Poggio, Isaac Chuang
This work argues that plateaus, sudden learning, and power-law scaling can all emerge from one simple mathematical model of how neural-network modes turn on.
Alex Shtoff
The Spectral Neuron replaces opaque nonlinearities with controllable eigenvalue-based functions, aiming to make expressive neural models more transparent and mathematically shapeable.
Junpeng Ren, Carlos Misael Madrid Padilla, Yanzhen Chen, Oscar Hernan Madrid Padilla
A new theory and method lets deep ReLU networks borrow strength across related datasets to improve nonparametric regression.
Runde Zha, Yuntian Gu, Chaohui Fan, Jia-lin Chen, Hai-Jun Liao, Tao Xiang
This work shows how several neural-network architectures can exactly encode exotic Motzkin quantum states with entanglement patterns that conventional tensor networks struggle to represent.
Yu-Hsi Chen, Abd-Krim Seghouane
PP-CPCANet replaces fragile covariance estimation with a robust, jointly learned orthogonal basis to improve domain-generalized representations.
Hyunmin Cho, Jaejun Yoo, Kyong Hwan Jin
A recurrent sinusoidal neural representation learns sharper images and 3D signals with fewer parameters and optimization steps.
Nicholas J. Cooper, François G. Meyer, Michael L. Roberts, Carlos Zapata-Carratalá, Lijun Chen, Danna Gurari
This paper gives a new theory for measuring how architecturally complex a neural network is and uses it to generate thousands of previously unexplored designs.
Paolo Baglioni, Christian Keup, Vincenzo Zimbardo, Rosalba Pacelli, Alessandro Vezzani, Raffaella Burioni, Pietro Rotondo
This paper develops a new theory for how wide Bayesian neural networks generalize, showing that finite-width effects can be captured with a renormalized kernel model that matches experiments reasonably well.
Connall Garrod, Jonathan P. Keating, Christos Thrampoulidis
This paper explains how depth alone can steer neural networks away from neural collapse and toward low-rank softmax-like solutions, revealing a new theoretical bias in training dynamics.
Shuo Huang, Lorenzo Fiorito, Lorenzo Rosasco, Tomaso Poggio
This paper shows that norm-constrained deep neural networks can learn sparse hierarchical functions efficiently, avoiding the curse of dimensionality by matching the structure of the target.
James Town, Etienne Boursier, Ben Lewis, Matthias Englert, Ranko Lazic
This paper proves that small two-layer ReLU networks can learn one neuron at a time in a saddle-to-saddle process, and it quantifies how close the resulting solution is to the simplest possible interpolator.
Balazs Szegedy
This paper shows that a function can be understood three equivalent ways: by how well it can be reconstructed from random samples, by a compact polynomial representation, or by a small neural network.
Shuang Chen, Juncai He, Xue-Cheng Tai
This paper builds a unified theory showing that neural flow models can approximate both functions and operators, connecting modern continuous-depth networks to residual and plain architectures.