Straight-Through Estimator

The Straight-Through Estimator (STE) is a widely used heuristic for Quantization-Aware Training (QAT), but its surrogate gradients can exhibit substantial mismatch with the underlying quantized objective, leading to noisy updates and parameter oscillations, particularly in ultra-low-bit regimes. We propose the Quantization-Aware Minimal-Norm Optimizer (Q-MINO), a temporal bundle method that combines gradient consensus, state-drift regularization, and an alignment constraint to construct stabilized, minimum-norm update directions from recent optimization states. Q-MINO solves the resulting constrained subproblem using a warm-started Frank--Wolfe procedure with a feasible fallback initialization. Theoretically, via a stochastic Lyapunov Kurdyka--Łojasiewicz (KL) framework, we show that Q-MINO achieves asymptotic neighborhood convergence. Moreover, we detail numerical experiments with Q-MINO at various quantizations.

Vector Quantization (VQ) is fundamental to discrete visual tokenizers that power modern autoregressive and masked image generation models. While recent shared-projection codebook methods have substantially advanced codebook utilization, training stability remains a critical and underexplored challenge. We argue that the root cause lies in the entanglement of the Encoder--Decoder and Codebook training: because neither module can reliably fulfill its own responsibility in isolation, the system can only function when the two subsystems happen to cooperate---a fragile condition that breaks down precisely when training is most stressed. We propose StableVQ, which revisits the proper learning objective of each module and resolves the problems that arise when each is trained to fulfill its own role independently. Concretely, (1) Dynamic STE corrects the instability in the Encoder's learning objective, enabling it to robustly optimize the reconstruction space under discrete regularization even when codebook utilization is low. (2) Region VQ Loss reconceives the Codebook's learning objective so that it can independently guarantee full tracking of the encoder output distribution, without relying on encoder oscillations to drive activation. (3) Decoupled Schedule recognizes that the distinct responsibilities of the Encoder--Decoder and the Codebook demand distinct optimization dynamics, and assigns each an independent learning rate schedule to ensure robust system-level behavior. Built on top of shared-projection codebooks, StableVQ is lightweight and introduces no learnable parameters. Experiments on ImageNet demonstrate consistent improvements in training stability, codebook utilization, and reconstruction quality across diverse codebook sizes and initialization settings.

We study the identity straight-through estimator (STE) for training a two-layer binary-activation network with hinge loss from the perspective of Statistical Learning Theory (SLT). Our central question is whether algorithmic stability can explain the statistical generalization of the estimator produced by the discontinuous STE training rule. In the saturated-output regime, the zero-initialized samplewise STE recursion is exactly the stochastic subgradient descent on the convex latent loss $(-yu^\top x)_+$. This representation makes a stability analysis possible. We derive an exact distance identity for two coupled updates and prove approximate non-expansiveness of the common-example map, with a quadratic defect only when the two latent margins straddle zero. We then obtain explicit $\ell_2$ on-average model-stability and generalization bounds, transferring stability isometrically from the latent vector to the full first-layer matrix. Combining stability with a standard optimization bound yields an explicit excess induced-risk guarantee and the rate $O(n^{-1/2})$ when $T=n^2$. Under margin separability, a complementary argument gives the optimal-order $O(R^2/(γ^2n))$ expected excess misclassification error for a randomized one-pass STE iterate and a corresponding majority-vote bound.

We introduce the propensity straight-through (PST) estimator for gradient-based learning through exact simulations of continuous-time Markov chains (CTMCs). These models describe discrete stochastic dynamics across physics, chemistry, and biology, but their categorical event choices prevent ordinary differentiation. PST preserves exact forward trajectories and differentiates, in the backward pass, the normalized reaction propensities the simulator already computes. The affine state update makes this rule exact for the one-step conditional-mean sensitivity. We derive the per-step discrepancy for one-step sensitivities composed along nonlinear trajectories in closed form, show that it vanishes for affine downstream dependence, and bound it quadratically in the stoichiometric jump size. Across reversible dimerization, a genetic oscillator, a 50-task repressilator suite, and experimental patch-clamp recordings, PST recovers parameters to within 0.06-1.7% and converges 2.4-3.0-fold faster than the Gumbel-Softmax straight-through estimator at matched computational cost. At deep-learning scale, PST trains all 203,796 parameters of a stochastic reaction network using exact discrete events from initialization. By differentiating an exact conditional mean, PST eliminates reaction-selection temperature tuning and provides a scalable route to learning through exact stochastic dynamics.

Training binary neural networks (BNNs) from scratch is dominated by the straight-through estimator (STE), whose forward/backward mismatch produces severe accuracy degradation as networks deepen. We study an orthogonal axis: when and where binarization is enforced during training. We introduce StoMPP (Stochastic Masked Partial Progressive Binarization), which gradually replaces clipped weights and activations with their hard binary counterparts layer by layer from input to output, using stochastic partial masks with soft refresh. StoMPP delivers two complementary benefits. As a standalone training rule, it provides a fully STE-free procedure that improves over vanilla STE with gains that grow with depth (ResNet-50 BNN: +18.0/+13.5/+3.8 on CIFAR-10/100/ImageNet), and the pattern holds across ResNet-18/34/50, MobileNetV2, and BERT fine-tuning. Composed with surrogate gradients by applying STE only to frozen entries, it reaches +27.1/+19.8/+17.7 over vanilla STE on the same setting. Underlying both regimes is a single mechanistic finding: progression order is decisive. Forward layerwise progression prevents depth collapse, reverse progression collapses to near-chance, and binary-weight networks (without binary activations) are insensitive to order. We trace this asymmetry to activation-induced gradient blockades: a committed binary activation severs gradient flow upstream, and ordering controls when these blockades form. To isolate the progression's contribution from any benefit conferred by STE, we conduct all ablations in the STE-free regime; the resulting characterization (schedule, refresh, ordering, dynamics) thus reflects the progression itself rather than its interaction with surrogate gradients.

Post-training quantization (PTQ) converts a trained full-precision model into low-bit weights without task-level retraining, while quantization-aware training (QAT) incorporates quantization into the training loop. Although PTQ is efficient and often accurate at moderate bitwidths, it can fail sharply at aggressive bitwidths; QAT is more expensive but can often recover the lost accuracy. We propose a unified geometric framework that explains both PTQ failure and QAT recovery. We model full-precision training as following a low-loss river inside a wider valley: a normal neighborhood of the river forms a nearly flat basin, while leaving this basin incurs a sharp loss increase. When the quantization grid is comparable to the basin width, local PTQ objectives, including rounding and Hessian-based second-order reconstruction, can select a high-loss deployed quantized point outside the basin even when nearby low-loss quantized points exist. In this regime, straight-through-estimator-based QAT has a useful bias: it evaluates gradients at the deployed quantized weights while updating latent full-precision weights, causing the gradient to sense the valley wall and acquire an inward component that steers subsequent quantized iterates back into the basin. We formalize this mechanism through a local landscape model, construct a geometric PTQ failure mode, and prove finite-time QAT recovery under local quantizer-compatibility assumptions. Experiments across vision and language models under multiple neural-network quantization schemes corroborate the predicted basin-crossing failure of PTQ and the corresponding recovery mechanism of QAT.

Stochastic kinetic models are ubiquitous in physics, yet inferring their parameters from experimental data remains challenging. For deterministic models, parameter inference often relies on gradients, which can be obtained efficiently through automatic differentiation (AD). However, AD cannot be applied directly to the Gillespie stochastic simulation algorithm (SSA), since sampling from a discrete set of reactions introduces non-differentiable operations. In this work, we adopt three gradient estimators from machine learning for the Gillespie SSA: the Gumbel-Softmax Straight-Through (GS-ST) estimator, the Score Function estimator, and the Alternative Path estimator. We use the estimators to evaluate gradients of steady-state and time-dependent observables, and compare their performance in representative biophysical systems with relaxation dynamics (bimolecular association) and oscillatory dynamics (repressilator). We find that the GS-ST estimator generally yields well-behaved gradient estimates, but exhibits diverging variance in challenging parameter regimes, which can cause parameter inference to fail. In these cases, other estimators provide more robust, lower variance gradients. Our results demonstrate that gradient-based parameter inference can be effectively combined with the Gillespie SSA, with different estimators offering complementary advantages.

Tree-based models are widely recognized for their interpretability and have proven effective in various application domains, particularly in high-stakes domains. However, learning decision trees (DTs) poses a significant challenge due to their combinatorial complexity and discrete, non-differentiable nature. As a result, traditional methods such as CART, which rely on greedy search procedures, remain the most widely used approaches. These methods make locally optimal decisions at each node, constraining the search space and often leading to suboptimal tree structures. Additionally, their demand for custom training methods precludes a seamless integration into modern machine learning (ML) approaches. In this thesis, we propose a novel method for learning hard, axis-aligned DTs through gradient descent. Our approach utilizes backpropagation with a straight-through operator on a dense DT representation, enabling the joint optimization of all tree parameters, thereby addressing the two primary limitations of traditional DT algorithms. First, gradient-based training is not constrained by the sequential selection of locally optimal splits but, instead, jointly optimizes all tree parameters. Second, by leveraging gradient descent for optimization, our approach seamlessly integrates into existing ML approaches e.g., for multimodal and reinforcement learning tasks, which inherently rely on gradient descent. These advancements allow us to achieve state-of-the-art results across multiple domains, including interpretable DTs rees for small tabular datasets, advanced models for complex tabular data, multimodal learning, and interpretable reinforcement learning without information loss. By bridging the gap between DTs and gradient-based optimization, our method significantly enhances the performance and applicability of tree-based models across various ML domains.

Automatic differentiation (AD) frameworks such as JAX and PyTorch have enabled gradient-based optimization for a wide range of scientific fields. Yet, many "hard" primitives in these libraries such as thresholding, Boolean logic, discrete indexing, and sorting operations yield zero or undefined gradients that are not useful for optimization. While numerous "soft" relaxations have been proposed that provide informative gradients, the respective implementations are fragmented across projects, making them difficult to combine and compare. This work introduces SoftJAX and SoftTorch, open-source, feature-complete libraries for soft differentiable programming. These libraries provide a variety of soft functions as drop-in replacements for their hard JAX and PyTorch counterparts. This includes (i) elementwise operators such as clip or abs, (ii) utility methods for manipulating Booleans and indices via fuzzy logic, (iii) axiswise operators such as sort or rank -- based on optimal transport or permutahedron projections, and (iv) offer full support for straight-through gradient estimation. Overall, SoftJAX and SoftTorch make the toolbox of soft relaxations easily accessible to differentiable programming, as demonstrated through benchmarking and a practical case study. Code is available at github.com/a-paulus/softjax and github.com/a-paulus/softtorch.

Machine learning models involving discrete latent variables require gradient estimators to facilitate backpropagation in a computationally efficient manner. The most recent addition to the Straight-Through family of estimators, ReinMax, can be viewed from a numerical ODE perspective as incorporating an approximation via Heun's method to reduce bias, but at the cost of high variance. In this work, we introduce the ReinMax-Rao and ReinMax-CV estimators which incorporate Rao-Blackwellisation and control variate techniques into ReinMax to reduce its variance. Our estimators demonstrate superior performance on training variational autoencoders with discrete latent spaces. Furthermore, we investigate the possibility of leveraging alternative numerical methods for constructing more accurate gradient approximations and present an alternative view of ReinMax from a simpler numerical integration perspective.

Exact stochastic simulation of continuous-time Markov chains (CTMCs) is essential when discreteness and noise drive system behavior, but the hard categorical event selection in Gillespie-type algorithms blocks gradient-based learning. We eliminate this constraint by decoupling forward simulation from backward differentiation, with hard categorical sampling generating exact trajectories and gradients propagating through a continuous massively-parallel Gumbel-Softmax straight-through surrogate. Our approach enables accurate optimization at parameter scales over four orders of magnitude beyond existing simulators. We validate for accuracy, scalability, and reliability on a reversible dimerization model (0.09% error), a genetic oscillator (1.2% error), a 203,796-parameter gene regulatory network achieving 98.4% MNIST accuracy (a prototypical deep-learning multilayer perceptron benchmark), and experimental patch-clamp recordings of ion channel gating (R^2 = 0.987) in the single-channel regime. Our GPU implementation delivers 1.9 billion steps per second, matching the scale of non-differentiable simulators. By making exact stochastic simulation massively parallel and autodiff-compatible, our results enable high-dimensional parameter inference and inverse design across systems biology, chemical kinetics, physics, and related CTMC-governed domains.