Paper says power analysis struggles to extract full model weights
Even a noiseless leakage adversary needs many traces, and deep-layer recovery suffers as models reduce entropy.
Stories per day · 30 days
Relatedweight-space-programming3expressivity2qwen2scientific-machine-learning2adversarial-attacks1adversarial-training1approximation-theory1arxiv1attention1brain-modeling1clinical-decision-networks1concept-based1constitutive-modeling1continuous-time1credit-assignment1decompilation1deep-learning1deep-ritz1
38 stories, newest first · one card per story
Stories tagged #neural-networks
ReLU networks with one hidden layer are weakly closed
One-hidden-layer ReLU networks are weakly closed enabling energy minimization in Deep Ritz and Sobolev training.
Weight-Space Programming enables localized neural edits with up to 29.45-point gains
16 trajectories on Qwen models reveal WSP's strong localized improvements but inconsistent generalization and reliability
AI forecasts grain growth under non-isothermal conditions
Recurrent neural networks conditioned on thermal data predict grain growth with mean relative errors below 5%.
Paper introduces uncertainty-guided surrogate for microfluidic droplet generator inverse design
Neural ensemble surrogates with targeted enrichment and PSO achieve nearly doubled accuracy for optimizing microdroplet devices.
HDPU Enables Neural Decisions at Sensor on FPGA
The HDPU processes frames on Artix-7 FPGA and spares the host all inference.
Language model audits and repairs flaws in small clinical decision networks
Claude Opus 5.5 audited and repaired flaws in radiotherapy networks with 433-4801 parameters
Show HN: Hacker Times, a New York Times-Inspired Version Hacker News
A new Hacker News site inspired by The New York Times aggregates AI research, tech updates and security stories.
1
LLMs are not superhuman at reading and writing neural network weights
A recent paper claims frontier LLMs can work with neural network weights directly, but this is not superhuman.
1
Vis-Lens generates preferred visual representations in neural networks and human brain using fMRI
Vis-Lens integrates a variational autoencoder and diffusion model to generate realistic preferred images by modifying activation-regularized latent codes.
Astra reads and writes neural network weights
Frontier LLM Astra identifies trained behaviors from raw weights and edits Qwen3.5-0.8B weights to improve math performance without gradients.
1
Neural decompilation candidate selection closes gap on obfuscated binaries
Generate-then-select pipeline with execution-guided selection reaches 88.8% pass@1 on obfuscated inputs.
Velocity Matching Objective Improves Dynamical System Learning
Simulation-free velocity matching approximates continuous-time evolution without rollout simulation or Jacobian constraints.
DeepWeave neuro-symbolic NIDS improves low-data detection and provides explanations
DeepWeave combines neural predictions with interpretable rules for robust, label-efficient NIDS.
Intermittent noisy gradients can leave learning variance unbounded
Average dynamics may still reach the correct optimum while learning variance grows without bound.
Concept-based attacks target neural-network logic, not class labels
Retraining on the attacks kept classical robustness gains while better removing semantic shortcuts.
Study measures how neural representations form during training
Path tortuosity, not endpoint displacement, predicted later drift and sensitivity to parameter perturbations.
Tutorial presents KKT-HardNet for constrained hybrid modeling
The framework enforces linear and nonlinear constraints via projection and releases code plus engineering benchmarks.
Paper urges four-dimensional treatment of learned fluid turbulence
Using direct numerical simulations larger than 28 TB per run, the authors find spatial and temporal checks can disagree.
SPCNNs approximate symmetric polyconvex energies by construction
The authors prove universal approximation without isotropy and illustrate stress regression plus a double-well envelope.
Partially trainable networks show a sharp approximation phase transition
For spherical biased-ReLU features at d=5 and τ=10, about 4.1% trainable width captures 90% of the bound reduction.
TWIG stores neural weights as shared paths of discrete decisions
On five LLM weight matrices, a flat power-of-two TWIG needs 9-15% fewer additions than GPTQ at equal error.
Paper derives attention architectures from geometric knowledge spaces
The framework recovers standard, causal, linear, and tensor attention plus recurrent networks, then proposes new geometric spaces.
Any ReLU network can be exactly simulated with two hidden layers
The authors give a homological-algebra proof using fillings and a separate direct combinatorial proof.