Differentiable reimplementations of the Atari VCS provide a complex, fully known ground-truth system for testing gradient-based explainable AI methods.
super hub Mixed citations
Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation
Mixed citation behavior. Most common role is method (62%).
210
Pith papers citing it
2,009
external citations · Pith
Method
62% of classified citations
abstract
Stochastic neurons and hard non-linearities can be useful for a number of reasons in deep learning models, but in many cases they pose a challenging problem: how to estimate the gradient of a loss function with respect to the input of such stochastic or non-smooth neurons? I.e., can we "back-propagate" through these stochastic neurons? We examine this question, existing approaches, and compare four families of solutions, applicable in different settings. One of them is the minimum variance unbiased gradient estimator for stochatic binary neurons (a special case of the REINFORCE algorithm). A second approach, introduced here, decomposes the operation of a binary stochastic neuron into a stochastic binary part and a smooth differentiable part, which approximates the expected effect of the pure stochatic binary neuron to first order. A third approach involves the injection of additive or multiplicative noise in a computational graph that is otherwise differentiable. A fourth approach heuristically copies the gradient with respect to the stochastic output directly as an estimator of the gradient with respect to the sigmoid argument (we call this the straight-through estimator). To explore a context where these estimators are useful, we consider a small-scale version of {\em conditional computation}, where sparse stochastic units form a distributed representation of gaters that can turn off in combinatorially many ways large chunks of the computation performed in the rest of the neural network. In this case, it is important that the gating units produce an actual 0 most of the time. The resulting sparsity can be potentially be exploited to greatly reduce the computational cost of large deep networks for which conditional computation would be useful.
hub tools
citation-role summary
method 16
background 10
citation-polarity summary
claims ledger
- abstract Stochastic neurons and hard non-linearities can be useful for a number of reasons in deep learning models, but in many cases they pose a challenging problem: how to estimate the gradient of a loss function with respect to the input of such stochastic or non-smooth neurons? I.e., can we "back-propagate" through these stochastic neurons? We examine this question, existing approaches, and compare four families of solutions, applicable in different settings. One of them is the minimum variance unbiased gradient estimator for stochatic binary neurons (a special case of the REINFORCE algorithm). A s
authors
co-cited works
representative citing papers
LCDD creates sparse carriers for SFT behaviors that SFT-Eraser can reverse, with ablations showing the sparse structure enables causal control.
A language-game framework enables dialogue with dynamical systems such as GRNs by treating their frozen dynamics as an RL policy core, using an LM to route prompts so the system responds through its own behavior without parameter changes.
INT4 quantization recovers up to 22 times more forgotten training data in unlearned LLMs, and the proposed DURABLEUN-SAF method is the first to maintain forgetting across BF16, INT8, and INT4 precisions.
A quantization vector derived from a donor model via weight-space arithmetic can be added to a receiver model to improve post-PTQ Top-1 accuracy by up to 60 points in 3-bit settings without receiver-side QAT or data.
A noisy top-k gated mixture-of-experts layer between LSTMs scales neural networks to 137B parameters with sub-linear compute, beating SOTA on language modeling and machine translation.
Gumbel-Softmax provides a continuous relaxation of categorical sampling that anneals to discrete samples for gradient-based optimization.
Gradient estimation of probabilistic programs reduces soundly to probabilistic inference after programmable coupling and factorization, enabling new low-variance estimators that beat baselines.
GB-FESO backpropagates a KL-divergence loss through a frozen conditional diffusion model's sampling trajectory to optimize system parameters so the generated ensemble matches a target free-energy surface.
FlexiSLM is the first spoken language model supporting dynamic and controllable frame rates on speech input and output, outperforming fixed-rate 7B models at high quality and enabling faster inference at lower rates like 6.25 Hz.
Task-specific LoRA adapters in continual learning exhibit significant low-rank subspace overlap, enabling LiteLoRA's learned gating to reduce active adapters by 20-70% while matching or exceeding prior performance.
StoMPP progressively binarizes BNN layers layerwise from input to output via stochastic masks, delivering depth-scalable accuracy gains in a fully STE-free regime by controlling activation-induced gradient blockades.
BiLoc is the first binary neural network framework for 6-DoF LiDAR pose estimation that uses an auxiliary objective to adaptively regulate information retention and achieve SOTA among BNNs on large outdoor datasets.
LogicIR is the first logic gate network for image restoration, using a UNet-like structure with differentiable bit decoding and index shuffling to deliver competitive results at lower computational cost.
Learning quantization-aware linear paths in weight space yields a midpoint whose direct quantization matches quantization-aware training performance without using straight-through estimators.
Placing trainable nonlinear functions on connections in analogue networks enables efficient representation of smooth continuous targets with hardware transfer at projected 30 microwatt power.
Timage generates text query overlays on images via Constrained Schrödinger Bridge to boost fine-grained spatial reasoning in vision-language models, outperforming larger systems on VMCBench with a 7B backbone.
CREST re-anchors global readouts in dynamics models to transient events via event-versus-rest contrast on learned features, reducing OOD error on gear, impact, and bearing systems while restoring event credit.
JOMP jointly optimizes mixed-precision quantization parameters and bit widths across neural video coding frameworks, achieving rate-distortion performance comparable to DCVC-FM while cutting bit operations by 87.6%.
Gumbel-BEARD automates Whisper layer selection with Gumbel-Softmax and BEST-RQ for self-supervised domain adaptation, matching fully supervised performance on 10h vs 133h data and setting new SOTA WERs on MyST and OGI datasets.
A geometric local landscape model explains PTQ basin-crossing failure at aggressive bitwidths and proves finite-time QAT recovery via straight-through estimator gradient bias under quantizer-compatibility assumptions.
MAVN adaptively selects and connects virtual nodes in MPNNs via learned dual-perspective preferences, proves it can realize any connectivity pattern, and reports up to 46.5% gains over backbones on nine datasets.
A U-Net surrogate with multigroup attention pooling is trained on OpenMC sensitivity data and combined with gradient optimization to generate grid-based critical experiment geometries that achieve c_k values up to 0.97757 for HALEU fuel validation.
Exposes a dynamic-probabilistic consistency gap in chaotic dynamical systems reconstruction and introduces the KAFFEE differentiable extended Kalman filter training framework to address it.
citing papers explorer
-
A Differentiable Atari VCS:A Complex, Fully Known Ground Truth for Explainable AI
Differentiable reimplementations of the Atari VCS provide a complex, fully known ground-truth system for testing gradient-based explainable AI methods.
-
Crafting Reversible SFT Behaviors in Large Language Models
LCDD creates sparse carriers for SFT behaviors that SFT-Eraser can reverse, with ablations showing the sparse structure enables causal control.
-
Language Game: Talking to Non-Human Systems
A language-game framework enables dialogue with dynamical systems such as GRNs by treating their frozen dynamics as an RL policy core, using an LM to route prompts so the system responds through its own behavior without parameter changes.
-
DurableUn: Quantization-Induced Recovery Attacks in Machine Unlearning
INT4 quantization recovers up to 22 times more forgotten training data in unlearned LLMs, and the proposed DURABLEUN-SAF method is the first to maintain forgetting across BF16, INT8, and INT4 precisions.
-
Zero-Shot Quantization via Weight-Space Arithmetic
A quantization vector derived from a donor model via weight-space arithmetic can be added to a receiver model to improve post-PTQ Top-1 accuracy by up to 60 points in 3-bit settings without receiver-side QAT or data.
-
Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer
A noisy top-k gated mixture-of-experts layer between LSTMs scales neural networks to 137B parameters with sub-linear compute, beating SOTA on language modeling and machine translation.
-
Categorical Reparameterization with Gumbel-Softmax
Gumbel-Softmax provides a continuous relaxation of categorical sampling that anneals to discrete samples for gradient-based optimization.
-
GradInf: Gradient Estimation as Probabilistic Inference
Gradient estimation of probabilistic programs reduces soundly to probabilistic inference after programmable coupling and factorization, enabling new low-variance estimators that beat baselines.
-
Gradient-Based Inverse Design of Free-Energy Landscapes with Diffusion Models
GB-FESO backpropagates a KL-divergence loss through a frozen conditional diffusion model's sampling trajectory to optimize system parameters so the generated ensemble matches a target free-energy surface.
-
FlexiSLM: A Dynamic and Controllable Frame Rate Spoken Language Model
FlexiSLM is the first spoken language model supporting dynamic and controllable frame rates on speech input and output, outperforming fixed-rate 7B models at high quality and enabling faster inference at lower rates like 6.25 Hz.
-
When One Adapter Speaks for Many: Discovering Low-Rank Redundancy in Continual Fine-Tuning
Task-specific LoRA adapters in continual learning exhibit significant low-rank subspace overlap, enabling LiteLoRA's learned gating to reduce active adapters by 20-70% while matching or exceeding prior performance.
-
Layerwise Progressive Freezing: A Training Scaffold for Depth-Scalable Binary Networks
StoMPP progressively binarizes BNN layers layerwise from input to output via stochastic masks, delivering depth-scalable accuracy gains in a fully STE-free regime by controlling activation-induced gradient blockades.
-
Learning 1-Bit LiDAR-based Localization with Auxiliary Objective
BiLoc is the first binary neural network framework for 6-DoF LiDAR pose estimation that uses an auxiliary objective to adaptively regulate information retention and achieve SOTA among BNNs on large outdoor datasets.
-
LogicIR: Logic Gate Networks for Image Restoration
LogicIR is the first logic gate network for image restoration, using a UNet-like structure with differentiable bit decoding and index shuffling to deliver competitive results at lower computational cost.
-
Neural Network Quantization by Learning Low-Loss Subspaces
Learning quantization-aware linear paths in weight space yields a midpoint whose direct quantization matches quantization-aware training performance without using straight-through estimators.
-
Low-power analogue neural networks with trainable nonlinear connections for continuous control
Placing trainable nonlinear functions on connections in analogue networks enables efficient representation of smooth continuous targets with hardware transfer at projected 30 microwatt power.
-
Timage: A Generative Text-in-Image Paradigm for Fine-Tuning Vision-Language Models
Timage generates text query overlays on images via Constrained Schrödinger Bridge to boost fine-grained spatial reasoning in vision-language models, outperforming larger systems on VMCBench with a 7B backbone.
-
When Dynamics Models Read the Wrong Time Steps: Label-Free Event Credit Re-Anchoring for Robust Global Readouts
CREST re-anchors global readouts in dynamics models to transient events via event-versus-rest contrast on learned features, reducing OOD error on gear, impact, and bearing systems while restoring event credit.
-
JOMP: Jointly-Optimized Mixed-Precision Quantization Across Neural Video Coding Frameworks and Buffering Strategies
JOMP jointly optimizes mixed-precision quantization parameters and bit widths across neural video coding frameworks, achieving rate-distortion performance comparable to DCVC-FM while cutting bit operations by 87.6%.
-
Gumbel-BEARD: Automatic Layer Selection for Self-Supervised Adaptation of Whisper in Low-Resource Domains
Gumbel-BEARD automates Whisper layer selection with Gumbel-Softmax and BEST-RQ for self-supervised domain adaptation, matching fully supervised performance on 10h vs 133h data and setting new SOTA WERs on MyST and OGI datasets.
-
Understanding Quantization-Aware Training: Gradients at Quantized Weights Bias to the Low-Loss Basin
A geometric local landscape model explains PTQ basin-crossing failure at aggressive bitwidths and proves finite-time QAT recovery via straight-through estimator gradient bias under quantizer-compatibility assumptions.
-
Learn When and Where to Connect: Adaptive Virtual Nodes for Dynamic Message Passing on Graphs
MAVN adaptively selects and connects virtual nodes in MPNNs via learned dual-perspective preferences, proves it can realize any connectivity pattern, and reports up to 46.5% gains over backbones on nine datasets.
-
Inverse Critical Experiment Design via Gradient Optimization and a Multigroup Attention-Based Neural Network Architecture
A U-Net surrogate with multigroup attention pooling is trained on OpenMC sensitivity data and combined with gradient optimization to generate grid-based critical experiment geometries that achieve c_k values up to 0.97757 for HALEU fuel validation.
-
The Dynamic-Probabilistic Consistency Gap in Chaotic Surrogate Modeling
Exposes a dynamic-probabilistic consistency gap in chaotic dynamical systems reconstruction and introduces the KAFFEE differentiable extended Kalman filter training framework to address it.
-
Binary Amplitude Modulation Suppresses Noise Up-Conversion in Coherent Diffractive Optical Networks
Binary amplitude masks in D2NNs deliver comparable MNIST/Fashion-MNIST accuracy to continuous modulation but superior noise robustness via an analytically derived metric C that is smaller for binary cases and computable from clean data.
-
A Minimal Bifurcation Model of Load Imbalance in a Softmax Mixture-of-Experts Router
A mean-field limit of a reinforcement-based softmax router for two experts shows a supercritical pitchfork bifurcation, with an external asymmetry unfolding it into a cusp of fold bifurcations.
-
RankE: End-to-End Post-Training for Discrete Text-to-Image Generation with Decoder Co-Evolution
RankE co-evolves AR policy and decoder via alternating ranking optimization, improving both FID and CLIP scores on LlamaGen-XL and Janus-Pro where policy-only RL degrades FID.
-
FTerViT: Fully Ternary Vision Transformer
FTerViT introduces fully ternary Vision Transformers with TernaryBitConv2d and TernaryLayerNorm operators, achieving 82.43% ImageNet top-1 at 6.09 MB with 15x compression.
-
Drifting Objectives for Refining Discrete Diffusion Language Models
TokenDrift refines discrete diffusion language models by applying anti-symmetric drifting to soft-token features during training, yielding large reductions in generation perplexity at low NFEs.
-
Not All Symbols Are Equal: Importance-Aware Constellation Design for Semantic Communication
A framework combining VQ-VAE, task-relevance scoring, DRL subset selection, and a learned semantic-aware constellation claims near-100% semantic protection probability and proves standard Gray-coded constellations are suboptimal under non-uniform importance.
-
Policy Optimization in Hybrid Discrete-Continuous Action Spaces via Mixed Gradients
HPO enables unbiased policy optimization in hybrid action spaces by mixing differentiable simulation gradients with score-function estimates, outperforming PPO as continuous dimensions increase.
-
All Circuits Lead to Rome: Rethinking Functional Anisotropy in Circuit and Sheaf Discovery for LLMs
LLM tasks are supported by multiple distinct circuits rather than unique mechanisms, demonstrated via Overlap-Aware Sheaf Repulsion and the Distributive Dense Circuit Hypothesis.
-
Quantum Parity Representations: Learnable Basis Discovery, Encoders, and Shadow Deployment
Hybrid quantum training discovers parity bases that improve accuracy 24-42% on binary tasks and recover performance on text benchmarks, with all inference remaining classical.
-
AffectCodec: Emotion-Preserving Neural Speech Codec for Expressive Speech Modeling
AffectCodec is an emotion-guided neural speech codec that preserves emotional cues during quantization while maintaining semantic fidelity and prosodic naturalness.
-
Revisiting Mixture Policies in Entropy-Regularized Actor-Critic
A new marginalized reparameterization estimator allows low-variance training of mixture policies in entropy-regularized actor-critic algorithms, matching or exceeding Gaussian policy performance in several continuous control benchmarks.
-
Approximation-Free Differentiable Oblique Decision Trees
DTSemNet gives an exact, invertible neural-network encoding of hard oblique decision trees that supports direct gradient training for both classification and regression without probabilistic softening or quantized estimators.
-
PhySPRING: Structure-Preserving Reduction of Physics-Informed Twins via GNN
PhySPRING uses differentiable GNNs to learn hierarchical coarsened spring-mass topologies and parameters from observations, delivering up to 2.3x speedup on PhysTwin benchmarks and comparable robot policy success rates in zero-shot Real2Sim substitution.
-
SplatWeaver: Learning to Allocate Gaussian Primitives for Generalizable Novel View Synthesis
SplatWeaver uses cardinality Gaussian experts and pixel-level routing to dynamically allocate varying numbers of Gaussian primitives for generalizable novel view synthesis.
-
Long Context Pre-Training with Lighthouse Attention
Lighthouse Attention enables faster long-context pre-training via gradient-free symmetrical hierarchical compression of QKV while preserving causality, followed by a short full-attention recovery that yields lower loss than standard full-attention training.
-
Non-Myopic Active Feature Acquisition via Pathwise Policy Gradients
NM-PPG optimizes non-myopic acquisition policies for costly features by enabling pathwise gradients via continuous relaxation and straight-through rollouts in POMDPs, outperforming SOTA baselines.
-
Learning to Theorize the World from Observation
NEO is a probabilistic neural model that induces compositional programs as a learned Language of Thought from non-textual observations and executes them via a shared transition model to enable explanation-driven generalization.
-
Mixed-Precision Information Bottlenecks for On-Device Trait-State Disentanglement in Bipolar Agitation Detection
MP-IB uses an 8x information asymmetry via FP16 trait heads and INT4 state heads to disentangle speaker identity from agitation in voice biomarkers, outperforming larger models on edge devices with low latency and suppressed identity leakage.
-
GETA-3DGS: Automatic Joint Structured Pruning and Quantization for 3D Gaussian Splatting
GETA-3DGS is the first automatic joint structured pruning and quantization framework for 3D Gaussian Splatting, achieving roughly 5x storage reduction on standard datasets without per-scene thresholds.
-
Model Compression with Exact Budget Constraints via Riemannian Manifolds
The budget constraint in discrete model compression defines a Riemannian manifold allowing exact-constraint first-order optimization via Riemannian Constrained Optimization (RCO) without extra hyperparameters.
-
GradMAP: Gradient-Based Multi-Agent Proximal Learning for Grid-Edge Flexibility
GradMAP enables fast offline training of fully decentralized neural policies for grid-edge flexibility by embedding a differentiable three-phase AC power-flow model and applying proximal surrogates in action space.
-
LearnPruner: Rethinking Attention-based Token Pruning in Vision Language Models
LearnPruner prunes vision tokens to 5.5% of the original count while retaining about 95% of VLM performance and delivering 3.2 times faster inference by fixing attention sink in encoders and using unbiased middle-layer attention in LLMs.
-
Beyond Static Collision Handling: Adaptive Semantic ID Learning for Multimodal Recommendation at Industrial Scale
AdaSID adaptively regulates semantic ID overlaps in multimodal recommendations to improve retrieval performance, codebook utilization, and downstream metrics like GMV.
-
Preserving Long-Tailed Expert Information in Mixture-of-Experts Tuning
A new SFT framework for MoE models combines bias-driven sparsification with gated condenser experts to retain long-tailed expert information, outperforming DenseMixer and ESFT by over 2.5% on math reasoning and commonsense QA benchmarks.
-
Relaxation-Informed Training of Neural Network Surrogate Models
Regularizers that penalize big-M constants, unstable neurons, and per-sample LP relaxation gaps during neural network training reduce MILP solve times by up to four orders of magnitude while preserving surrogate accuracy.
-
AAC: Admissible-by-Architecture Differentiable Landmark Compression for ALT
AAC is an admissible-by-architecture differentiable compressor for ALT landmarks that achieves near-optimal coverage on road networks with zero admissibility violations and faster median queries than FPS-ALT.