Overview

Backpropagation uses the chain rule to send error signals from a model’s outputs backward through its layers, providing gradients for updating parameters. Work improves its efficiency and reach through memory-saving gradient methods, parallel scans, gradient approximations, and better scheduling, including for recurrent models and large language models. Alternatives such as feedback alignment, predictive coding, and forward-only local learning aim to reduce global backward communication or support biological and physical hardware. The central tension is that these alternatives may improve locality or memory use, but often struggle to match backpropagation’s accuracy, convergence, or measured end-to-end energy efficiency.

Latest papers

48 papers6 papers in the last four weeks, up 20% on the four weeks before. 0.0% of all new papers.

Nov 9, 2026

A method to automatically discover symbolic local learning rules

We present a general method for discovering symbolic local learning rules that can replace backpropagation, applicable, in principle, to any deep learning architecture and task. The method automatically discovers local loss functions specific to each layer by (1) constructing a dictionary of monomials of local tensor variables, inspired by Penrose diagram notation, (2) regressing over coefficients weighting these monomials to build layerwise polynomial losses whose gradients together minimize a target global loss. We test the method on classification, reconstruction, and self-supervised learning tasks with small multilayer networks and minimal datasets. Across all tasks, the discovered rules induce non-trivial interactions between layers, which lead to significant reductions in the corresponding global losses. Furthermore, the method selects sparse combinations of monomials, finding local rules that are more interpretable than those generated by MLP-based alternatives. The potential implications of our method range from energy-efficient deep learning to the discovery of biologically plausible learning mechanisms.
Nov 3, 2026

CAPSkip: Cancellation-Aware Prefix Skipping for Selective Backpropagation

Per-step selective back-propagation forwards every example of a mini-batch but back-propagates only some of them. Most rules rank examples by loss or gradient magnitude, but such scalar scores ignore cancellation between per-example gradients: the gradient removed by a set can be much smaller than its individual magnitudes suggest. We introduce CAPSkip (Cancellation-Aware Prefix Skipping). Among the examples the model already classifies correctly, CAPSkip skips the longest prefix, in increasing order of last-layer gradient norm, whose vector sum stays within a relative budget bb of the batch's last-layer gradient. In exact arithmetic this bounds the change of the raw final-layer gradient, using rank-one contributions already available from the forward pass; the bound does not extend to deeper layers, optimizer updates, or the training trajectory. Because the cumulative distortion is not monotone in the prefix length, CAPSkip scans all prefixes; a factorised Gram matrix makes the scan cheap and keeps it on the device. With activation reuse and selected-row backward, CAPSkip fine-tunes DeiT-S on CIFAR-100 for 103 epochs in 0.576x the time of 103-epoch dense training, and is +0.25 pp [+0.08, +0.42] above an approximately time-matched 59-epoch dense run (n=3).
Nov 2, 2026

Which Chains to Calibrate for Physical Backpropagation: Per-User Factors Versus Base-Station Chains

In physical backpropagation, users return loss derivatives over the time-division duplex uplink, giving the base station the downlink gradient without channel reconstruction. Reciprocity calibration conventionally prioritizes base-station chains, as column-normalized zero-forcing reduces a per-user factor to a stream rotation. This priority can reverse. With ideal base-station chains and no noise, per-user amplitude errors make the returned field a weighted sum-rate gradient. With zero-forcing full service at both optima, the loss is independent of array size. Otherwise the selection penalty, serving the wrong users, carries 86 to 94 percent of the amplitude-induced loss in five of six configurations. Per-user phase errors mainly slow the iteration. Base-station chain errors scale the returned field per antenna, and at 3 dB and 25 degrees they cost one user 1.3 dB of signal-to-noise ratio (SNR) that no array size removes. With 80 users, such errors on both sides, and mismatched pilots, oracle per-user and base-station calibration leave 6.2 and 12.3 percent loss at 256 antennas on clustered channels. This ordering holds at 30 dB SNR and reverses for 1 dB, 10 degree user errors. A capped one-echo estimator comes within 2.4 percentage points of oracle per-user calibration but underperforms no calibration at 10 dB SNR.
Nov 1, 2026

Detailed Derivation of The Backpropagation Algorithm

Backpropagation is the standard method for computing the gradients that train neural networks, but it is usually presented in a compact vector form that skips the intermediate steps. This tutorial derives backpropagation step by step for a fully connected feedforward network with sigmoid activations and a squared-error loss. Using a running example with three layers, we first differentiate the error with respect to a single weight by following its path through the network with the chain rule. We then show how summing over all paths leads to the error terms δδ and to the four general backpropagation equations in matrix form. Every term is annotated with its dimensions, so each equation can be checked for consistency. The derivation is then extended to mm training examples, where the weight gradients become a single matrix product and the bias term corresponds to broadcasting. It is also extended to mini-batch gradient descent, including why the gradient of a mini-batch estimates the full gradient. Batch and mini-batch training are both given as algorithms. Finally, each equation is linked to the corresponding line of a NumPy implementation trained on the AHDBase dataset of handwritten Arabic digits, and the complete code is included.
Oct 23, 2026

Neural LoFi with Backward Coupling: A Spectral Theory of Cross-Layer Feature Learning

During joint training, backpropagation allows later layers to reshape earlier representations, a feedback mechanism absent from purely layer-wise learning. We extend Neural Low-Degree Filtering (Neural LoFi), a recently proposed layer-wise spectral surrogate for feature learning, to capture this interaction. Starting from the exact backpropagation gradients, we derive a local coupling between a neuron's incoming weights and its corresponding next-layer column. In a controlled small-initialization regime with smooth activations and fixed surroundings, these dynamics amplify singular modes of a cross-layer correlation operator. Each growing mode pairs an earlier feature with an upper-layer response through shared neuron coefficients. This motivates a backward-coupling algorithm in which upper-layer sensitivities generate sample-dependent effective labels for earlier spectral steps, while corrected features refine upward transmission. A function-space variational principle characterizes the selected feature-response pairs by maximizing the label correlation of their product under kernel-norm constraints. For three-layer networks, we derive self-consistent continuous-time dynamics for the filtered hidden-layer features, the activation kernels, and the network output under the proposed backward coupling. In a hierarchical multi-index model, we establish a feature-recovery separation: backward coupling recovers lower-layer directions invisible to the original forward label-weighted statistic. The linear specialization recovers residual-dependent singular-mode dynamics, while paired modes provide a dynamical mechanism for linked low-rank covariance components consistent with the rainbow-network picture. Experiments with fully connected and convolutional networks on binary CIFAR-10 show improved accuracy over forward Neural LoFi at matched widths.
Oct 20, 2026

How Much Gradient Does Truncated Backprop Lose in Looped Transformers?

Looped transformers reuse one block of layers many times and are often trained by backpropagating through only the last k loops. We measure how much of the weight gradient this throws away, exactly and for every k from one backward pass, on nine looped language models, and report three findings. First, on a trained model this lost gradient is set near the output: once the loop count is well above k, it barely changes with the loop count (within about 0.03 from 32 to 64 loops on the three models we tested, as a pre-registered test predicted for Parcae). Where the backward signal shrinks, ρkρ^k, with ρρ the signal's fitted shrink per loop, approximates it at moderate loop counts but overstates it at long ones, because the fitted ρρ drifts up as the fit takes in more loops. On Ouro the signal grows, and keeping two of four loops still loses 81% of the final output's gradient. Second, the lost gradient falls during fine-tuning, by 0.07-0.22 within 300 steps on three of four models, on held-out problems too. The fall is not just the read-out's share of the gradient; training only the layers outside the loop reproduced up to half of it, and it also falls under truncated training. So a trained model's value overstates what truncation removes later. Third, we could not show that it predicts the training cost of truncation better than rules that need no measurement: on two new models it tied Parcae's cost at the same k, in a test too small to separate them.
Oct 12, 2026

FidLIF: Parallel Training of High-Fidelity Spiking Neurons with Exact BPTT

Although spiking neural networks (SNNs) show promise in energy-efficient computing through event-driven computation and stateful temporal processing, their recurrent membrane dynamics render their training inherently sequential and inefficient for long sequences. Existing methods achieve full-sequence parallelism in SNNs through two main strategies: redesigning or approximating spiking dynamics to avoid reset-coupled temporal dependencies, or relying on iterative relaxed fixed-point solvers that suffer from excessive memory costs and iteration-dependent backward overhead. To address these limitations, we present FidLIF, a novel framework that reconciles the incremental update and hard-reset behavior of standard spiking LIF neurons with parallel forward propagation and BPTT-equivalent backward propagation. FidLIF iteratively recovers the full-sequence membrane-potential trajectory in parallel. By exploiting the lower bidiagonal structure induced by the local recurrence of LIF neurons, FidLIF formulates each forward trajectory correction and backward adjoint propagation as associative scans with an O(log⁡T)O(\log T) temporal depth. This formulation also supports input-dependent neuronal parameters and allows the same neuron to perform full-sequence parallel training and incremental step-by-step inference at deployment. In our numerical convergence evaluation, FidLIF achieves a mean absolute error below 10−710^{-7} relative to the matched serial relaxed LIF dynamics, while yielding gradients numerically equivalent to exact BPTT. Our Triton implementation delivers favorable long-sequence latency and memory scaling, while achieving competitive performance on four sequential classification benchmarks. By reconciling parallel training with step-by-step execution, FidLIF enables high-fidelity spiking dynamics to benefit from parallel hardware without compromising incremental inference at deployment.
Oct 8, 2026

Connected Self Forcing: Beyond Local Learning in Video Autoregression

To stream long videos while maintaining visual quality and temporal consistency, Self Forcing mitigates exposure bias through self-rollout training on self-generated histories with key-value (KV) caching. To keep memory manageable, it detaches historical caches, preserving forward dependencies between chunks but severing the backward gradient paths. We introduce Connected Self Forcing, a training framework that reconnects gradient paths across autoregressive chunks, allowing feedback from later predictions to guide how earlier context is generated. These connections go beyond historical KV-writing: gradients pass through generated latents into the computations that produced them, linking the generation of earlier context to its use in later predictions. To make this connected training memory-efficient, we develop shortcut gradient replay, which recovers cross-chunk gradients without retaining the full rollout computation graph. Integrated with distribution matching distillation, Connected Self Forcing trains historical chunks according to both their direct supervision and their contribution to subsequent generation. Experiments on autoregressive video generation show improvements in long-horizon visual quality and temporal consistency, without changing the inference procedure.
Sep 30, 2026

Backward-State Policy Is Part of the Learning Algorithm

Low-precision training rounds tensors that the backward pass reads again, often for several gradients; each use can read the forward's rounded value, the original, or a new random rounding. This backward-state policy looks like a memory and precision detail, settled by copy accuracy and final loss. We argue that it is part of the learning algorithm, and that neither check shows whether it is right. Copy accuracy does not decide the outcome: in three pairs of 390M runs with an emulated FP8 backward, training fails when attention's backward reuses the forward's rounded output and succeeds with a new rounding from the same distribution. Even the most accurate copy, the original itself, can be wrong by our reference: the gradient of the forward pass as it actually ran, with gradients passed through rounding unchanged. For example, a normalization output stored in low precision feeds two gradients: the gain's gradient needs the original, but the next layer's weight gradient needs the rounded value that layer multiplied. Final loss, the other check, does not rule out the error of reading the original for both: it persists in models trained with such a store, while planned loss comparisons stay within a margin fixed in advance. We therefore derive from this reference which value each use must read, or which substitute gives the same gradient on average with the forward held fixed, and check these per-use requirements on single operators, without training. In three tests using PyTorch and Transformer Engine, the requirements predicted beforehand whether reuse changes what the backward computes on average relative to an independent copy, and every prediction held. Backward-state policy is thus part of the learning algorithm: it should be specified and checked use by use, not settled by copy accuracy and final loss.
Sep 27, 2026

The Price of Locality: Why Forward-Forward Underperforms Backpropagation?

The Forward-Forward Algorithm (FFA) replaces backpropagation (BP) with layer-wise local contrastive objectives, eliminating the backward pass and the need to retain intermediate activations, yet suffers a persistent performance gap with BP that worsens with depth. This paper diagnoses two structural failure modes: an optimization floor arising from concurrent local updates; and a geometric collapse of layer representations driven by the local update mechanism. On the optimization side, we prove that the FFA loss satisfies the Polyak--Lojasiewicz inequality at each layer; however, simultaneous layer updates induce inter-layer representation-distribution drift, so each layer optimizes against a moving input distribution and incurs an error floor. On the representational side, the pairwise similarity kernel of layer representations contracts exponentially toward rank one as depth increases, collapsing the diversity of per-layer error signals. This collapse bounds FFA's effective learning capacity, which measures the diversity of gradient information across layers, independently of depth, whereas BP's chain-rule signal preserves per-layer diversity, yielding a capacity that scales with depth.
Sep 26, 2026

LocalProp: Neuro-Localized Memory-Efficient Backpropagation

The current deep learning training paradigm employs end-to-end backpropagation, regardless of the training stage, i.e. pre-training or fine-tuning. However, backpropagating through the entire model is neither biologically plausible nor memory efficient, since learning inside the brain is highly localized. Therefore, we propose LocalProp, a training procedure that locally updates the weights of a model. Our neuro-localized weight updates follow the "pre-training then fine-tuning" paradigm, where the pre-training is based on I-JEPA. After locally updating the weights, a pruning operation is performed, followed by a short final fine-tuning phase. Pruning helps by sending the learning signal from higher blocks to lower blocks. We perform experiments on several datasets, including large-scale benchmarks such as ImageNet, and empirically show that LocalProp reaches good performance at a fraction of GPU peak memory. By varying the number of jointly optimized blocks, we identify gradient-propagation span as a practical control over the accuracy-memory trade-off.
Sep 18, 2026

The Ups and Downs of Backprop Weights

Backpropagation (BP) has driven the remarkable success of modern deep learning by enabling large hierarchical networks to learn complex functions end-to-end. Yet it does not by itself determine how parameters should be organized so that functional components can be reused and adapted selectively. For example, object recognition and motion prediction may depend on overlapping parameter sets, making them difficult to isolate or modify independently. We call this condition weight entanglement. Modern architectures dynamically select which parts of a network process each sample: nonlinearities gate units, attention selects interactions, and Mixture-of-Experts architectures route inputs to modules. Yet such selection does not ensure that the same functional component remains linked to an identifiable parameter set across samples. We propose weight operators: parameterized modules that implement reusable functional components and can be composed at inference to form the function required by each sample. Learning proceeds in two stages: the model first infers the required operator composition, then updates only the selected operators' parameter sets. Vector Networks (VNs) provide one implementation. They couple operator selection to local error-driven updates within each layer and show that learned operators can be reused in combinations absent from training while updates remain restricted to the selected parameter sets. This provides a basis for testing functional parameter identifiability: whether an operator remains linked to the same functional component during learning. We argue that functional parameter identifiability may provide an organizing principle for models that systematically reuse and recombine learned functions while adapting only the components that need to change.
Sep 11, 2026

Large Distant Gradients Need Not Be Reliable: reliability-weighted credit assignment for long-horizon autoregressive forecasting

In autoregressive forecasting, long prediction rollouts provide distant supervision, but backpropagation through time (BPTT) carries gradients from those losses through many autoregressive steps. Repeated Jacobian products can make distant gradients dominate the update while amplifying predictable signal and unpredictable innovation together; a large distant gradient therefore need not carry reliable learning signal. Motivated by this, we introduce Internal Dual-Wiener routing (Internal-DW), a backward-only intervention that preserves the full forward rollout and all step losses while reliability-weighting internal gradient routes. At each residual block, we derive bounded Wiener gains for the identity and nonlinear routes that balance preserving predictable learning signal against suppressing unpredictable variation, and estimate them from route-level gradient statistics and an explicit noise model. In a controlled system with known gradient signal-to-noise ratio (SNR), we show that distant gradients can grow even as their SNR falls, and that Internal-DW reduces error in recovering predictable gradient signals and improves forecasting. On four history-dominated, weak-drive testbeds, Internal-DW reduces forecast error by 5.2%-13.8% relative to full BPTT, outperforms gradient clipping and Jacobian regularization on three testbeds, with similar performance on shear flow, and outperforms validation-selected truncated BPTT (TBPTT) on three. It also extends or preserves the fitted optimal training-horizon range across these four testbeds. Across benchmarks, the current Internal-DW estimator has a clear applicability boundary: its benefit diminishes or reverses when usable history is limited or when the selected sampler fails to represent dominant drive-dependent variation. The results show that retaining long-horizon supervision does not require trusting every backward contribution equally.
Sep 8, 2026

Learning Length-Extrapolatable Recurrent Models

Recurrent models provide a natural path to long-context modeling, yet models trained with backpropagation through time (BPTT) often fail beyond their training horizon. Classical analyses emphasize gradients that vanish or explode along temporal paths. However, dense per-token losses can still train a shared recurrent rule despite severe decay, showing that decay alone does not determine whether learning fails. We instead study state credit: the signal through which future losses reach earlier recurrent states before contributing to parameter updates. Accordingly, we intervene directly on state credit and propose Credit Stabilization through Time (CST). During backward propagation, CST locally rescales the state-credit signal to stabilize its norm without rotating the component being corrected, while leaving the forward computation unchanged. Because controlled synthetic tasks and real data exhibit different credit dynamics, we specialize CST to each regime. In both settings, CST improves performance beyond the training horizon, with gains observed at up to 128x the training length.
Sep 1, 2026

Wave Function Backpropagation with Explicit Temporal-Interval Dynamics

Conventional neural networks learn predominantly through affine transformations followed by nonlinear activations, while elapsed time is often treated as an auxiliary feature or assumed to be uniformly sampled. This paper introduces Wave Function Backpropagation (WFB), a wave-parameterized learning formulation in which neural responses are represented by learnable amplitude, wavenumber, angular frequency, and phase. The formulation associates an observed state with its temporal interval Delta t through the phase of a differentiable spatiotemporal wave. We derive standard WFB gradients and a spatial-curvature correction based on the Laplacian of the wave response. WFB is instantiated in a deliberately feed-forward trajectory predictor to provide a controlled proof of concept; sequence learning is outside the scope of the present evaluation. With motion features, STD-WFB using real intervals reduces average displacement error (ADE) by 20.4% relative to the original FFN baseline. In a new position-only evaluation that removes temporal leakage through precomputed velocity and acceleration, real-interval WFB reduces ADE by 10.4% relative to the original FFN and remains competitive with parameter-matched ReLU controls, obtaining 2.1% lower mean ADE than the matched FFN with explicit Delta t. Shuffled-interval WFB attains the lowest mean ADE, indicating that the present evidence supports the effectiveness of the wave representation but does not attribute the gain to interval alignment. These results establish WFB as a viable structured feed-forward learning formulation and define a clear basis for subsequent architectural studies.
Aug 20, 2026

The Forward-Backward Disconnect: State Dynamics, Credit Assignment, and Biological Grounding in Neural Computation

A recurring pattern in neural computation is the reintroduction of dynamical and biological structure into models originally simplified for scalable optimization. Early feedforward networks reduced biological neurons to threshold or rate-like summation units, an abstraction compatible with global-gradient training at scale. Since then, forward computation has diversified: modern architectures carry recurrent state, retrieve from long contexts through attention and associative memory, compress histories through structured state-space dynamics, evolve in continuous time, settle to implicit equilibria, and communicate through sparse spikes. Training has diversified less. Scalable learning remains concentrated around backpropagation, backpropagation through time, adjoint methods, implicit differentiation, and surrogate-gradient variants. We call this asymmetry the forward-backward disconnect and develop a taxonomy spanning neural model families along three coupled axes: state-dynamics structure, credit-assignment mechanism, and biological grounding. Forward and learning grounding are treated separately, and the unit of analysis is the architecture-learning configuration rather than the architecture name alone. Across static, recurrent, attention-based, state-space, continuous-time, implicit, spiking, biologically plausible, and neuromorphic families, forward dynamics have diversified while the highest demonstrated scales remain concentrated in global or closely gradient-derived error-propagation mechanisms. Closing this disconnect requires better alignment among state dynamics, credit assignment, and computational substrate.
Aug 19, 2026

ERASE: EaRly bAckpropagation SchEdule for Faster Training of Modern Recommendation Systems

Lightweight proxy models enable rapid experimentation without repeatedly training frontier-scale systems, but their small kernels often leave modern accelerators underutilized. Conventional training compounds this inefficiency by scheduling the forward and backward passes as disjoint phases, so spare capacity in one cannot be filled by work from the other. We reinterpret the detachment mechanism of Forward-Forward (FF) as a scheduling primitive: given a local objective, detaching a block's output removes downstream gradient dependencies, making its backward pass ready when its forward pass finishes. ERASE launches each detached subgraph's backward pass early on a separate CUDA stream, overlapping it with subsequent forward work. Execution trace on a lightweight transformer demonstrates this overlap and its limit: a kernel that saturates the device leaves no capacity for concurrency. On a large-scale click-through-rate model, detaching six dense subarchitectures improves training throughput by up to 9.51%9.51\% while keeping normalized entropy close to the baseline.
Aug 12, 2026

Unifying Physical Backpropagation

Physical computing systems exploit device dynamics for computation, but their gradient-based optimization is challenging: backpropagation through a digital twin suffers from a model-reality gap. On-device gradient computation could resolve this issue, and a handful of theoretical and experimental studies have proposed ways to achieve it. Yet a unifying theory identifying when a physical system can compute the gradient of its own performance has been missing. Here we develop such a unification based on the adjoint method: we identify sufficient conditions under which the adjoint field required for formally exact gradients can be generated on the same hardware that performs the computation. Linear and nonlinear systems obey fundamentally different conditions: for linear systems, damping or gain is admissible provided reciprocity is preserved. For nonlinear trajectory systems, the sufficient conditions are reciprocity of the linearized system and the existence of a time-reversal mirror. Algorithmically, the nonlinear case requires infinitesimal nudging, whereas linear systems admit a finite-amplitude experiment. We recover (quantum) Equilibrium Propagation, Hamiltonian echo backpropagation, fully forward mode training and in situ gradient methods in integrated-photonic and free-space-optical systems. Finally, we show that reciprocity is a special case of more general intertwining conditions. For linear systems, these permit exact on-device gradients in a class of non-Hermitian, non-reciprocal systems. For nonlinear trajectories, they combine with generalized time-reversal mirrors to cover, e.g., PT-symmetric equations. The framework also includes time-dependent parameters and Onsager-reciprocal dynamics, providing a unified basis for formally exact physical learning.
Aug 7, 2026

Learning in Deep Networks under Dale's Constraint

Biologically plausible learning models aim to explain how neural circuits can implement effective learning under the constraints of real neurons. Although significant progress has been made, a major remaining challenge is that existing models often allow neurons or synapses to represent mixed-sign values, both positive and negative, in violation of a basic aspect of cortical circuitry -- Dale's constraint: biological neurons are either excitatory or inhibitory, but not both, and synapses cannot change sign. In this work, we address this discrepancy by introducing a biologically motivated neural architecture in which both neural activations and learning signals are represented by non-negative activity, and synapses have fixed sign, while still supporting backpropagation-like learning. Our approach uses two complementary interacting non-negative channels to represent positive and negative contributions, inspired by evidence of on-off representations in the brain. These channels are implemented through a simple neural circuit motif, which is repeated throughout the network in both bottom-up and top-down pathways. Combined with a local Hebbian learning rule, the resulting model propagates learning signals and updates weights using only local interactions between neurons. We show theoretically that our learning scheme can exactly recover the backpropagation update despite relying solely on non-negative error signals. Empirically, beyond satisfying stronger biological constraints, the on-off architecture learns efficient representations, yielding substantial gains over comparable vanilla networks on the Tiny ImageNet benchmark. These results demonstrate that effective learning can emerge from biologically plausible mechanisms without requiring mixed-sign signals, providing a step toward more realistic models of neural computation.
Jul 29, 2026

Reconstructing Backpropagation from Forward Fluctuations in Noise-modulated Neural Networks

A Noise-modulated Neural Network (NNN) learns and infers only in the presence of noise, treating noise as a computational resource rather than a disturbance. The noise lets it learn efficiently by backpropagation while transmitting spike-like signals, but backpropagation needs a reverse path through transposed weights, the weight transport problem, which undermines biological and neuromorphic plausibility. Forward-only alternatives typically substitute a different objective or fixed random feedback, sacrificing stability and accuracy. We show that backpropagation itself can be reconstructed in the NNN from forward-pass statistics alone: a weight mirror estimates each weight matrix from the covariance between a previous-layer unit's output and the next-layer unit's input, and combining it with local differential estimation inside the units propagates the output error recursively along the computational graph, with no transposed-weight readout and no backward data path. The resulting gradient is empirically near-unbiased, and with local per-weight Adam updates it matches the final accuracy of backpropagation on simple regression tasks. With uniformly distributed noise, the local operations reduce to polynomials and comparators, making the whole system, learning rule included, well suited to digital circuits. Thus, in the NNN, noise is a resource not only for inference but also for reconstructing backpropagation.
Jul 27, 2026

Continual-RL for Generalization in Autonomous Racing on the RoboRacer Platform

A key challenge in modern robotics is to adapt to changing environments, a challenge that is exacerbated when simulations cannot encompass every possible real-world configuration, and therefore Reinforcement Learning (RL) in the physical world becomes necessary. Continual Reinforcement Learning provides the tools to address this challenge; however, both the frameworks and the methods remain underexplored. Autonomous Racing and in particular the RoboRacer competition provide a testing ground for such methods, as learning to drive on a new track-floor combination with the least amount of new experience naturally frames a continual learning problem. This work tries to address this gap by proposing a continual RL framework based on Continual Backpropagation that is able, with only real-world data, to train a generalistic policy on a set of tracks and then fine- tune it within 15 minutes to outperform classical controllers. Furthermore, a comparison method based on offline RL is proposed, and a simulation analysis of the plasticity properties of the methods is conducted.
Jul 20, 2026

JAGG: Jacobian-Aggregated Group Gradient for Efficient GRPO Training of Diffusion Models

Group Relative Policy Optimization (GRPO) is a powerful reinforcement learning algorithm for aligning generative models with human preferences. While successful in large language models~\cite{shao2024deepseekmathpushinglimitsmathematical}, its extension to diffusion and flow matching models introduces a severe computational bottleneck: gradients must be back-propagated through the high-capacity DiT backbone at every timestep of the sampling trajectory, making high-resolution text-to-image (T2I) training prohibitively expensive. Training-free DiT inference acceleration methods (e.g., ΔΔ-DiT, ScalingCache) exploit the fact that DiT hidden states and velocity predictions vary smoothly and nearly linearly along the trajectory. We ask whether the same linearity can reduce the backward-pass cost of DiT RL training, and answer affirmatively with JAGG (Jacobian-Aggregated Group Gradient), which reduces full transformer backward passes from WW to 22 per group of WW consecutive steps. JAGG approximates intermediate-step Jacobians via tt-weighted interpolation of the endpoint Jacobians, then aggregates per-step upstream signals into two composite gradients applied through a single joint backward pass. We prove this interpolation is exact when the velocity is linear in (z,t)(z,t), and a cosine-similarity routing rule (jagg\_frac) deploys JAGG only where the assumption holds. Experiments on T2I benchmarks show JAGG delivers ∼\sim2×\times backward speedup with negligible quality degradation. The code for this work can be accessed through https://github.com/SchumiDing/JAGG.
Jul 18, 2026

Building a Neural Network from Scratch: Implementation, Evaluation, and Optimization

The widespread adoption of high-level deep learning libraries, while accelerating model development, has increasingly abstracted away the internal mechanics of neural networks, creating a gap between practical usage and fundamental understanding. To address this, the paper presents a self-contained neural network framework implemented entirely from scratch without relying on automatic differentiation or pre-built deep learning modules. The implementation encompasses all essential components, including multi-layer architectures, diverse activation functions, regularization techniques, and state-of-the-art optimizers. Beyond serving as a pedagogical instrument that demystifies forward/backward propagation, gradient dynamics, and optimization landscapes, the framework demonstrates robust performance when applied to a multi-class classification task, successfully validating its correctness, numerical stability, and generalization across varied configurations. The extensible design and clean modularity further position it as a reliable baseline for educational purposes and future research exploration.
Jul 18, 2026

Backpropagation-Free Trunk Training via the Split Forward Gradients

Backpropagation makes training deep networks memory intensive because it must store intermediate activations. Forward-mode methods avoid this cost, but their gradient estimates become increasingly noisy as the number of trained parameters grows. We introduce Split Forward Gradient (Split-FG), which splits a network at an intermediate representation: it computes the output head gradient exactly and estimates only the trunk gradient with a Jacobian--vector product. This reduces estimator variance and requires no backward pass through the trunk, while retaining an Adam-style convergence guarantee. Our experiments reveal an important practical failure mode. On WikiText-103, naive forward-gradient training of the trunk performs worse than leaving a randomly initialized trunk frozen, likely because Adam updates every noisy, under-determined trunk coordinate too aggressively. Simply using a much smaller learning rate for the trunk reverses this result: a 1616M-parameter GPT-2-style model reaches validation perplexity 387387, compared with 668668 for the frozen-trunk control and 2,8852{,}885 for a matched pure forward-gradient baseline (backpropagation reaches 150150). Split-FG also produces the strongest backprop-free results on our tabular benchmarks and reaches 60.5%60.5\% on CIFAR-10 and 35.2%35.2\% on CIFAR-100 with a heavy-head design. It reduces peak memory by up to 35%35\% relative to matched backpropagation, although the performance gap widens as the forward-mode trunk grows.
Jul 13, 2026

Backpropagation as a Nilpotent Linear System

Backpropagation is the computational engine of deep learning, yet its mathematical structure is typically treated as a procedural traversal of computational graphs. We present a global operator theory of the F-adjoint framework, which reformulates the layerwise backward recursion of an LL-depth feedforward network into a single linear system (I−\cB)\Xs=\bG(I-\cB)\Xs=\bG, where \bG\bG is a source vector. We prove that the global backward operator \cB\cB is strictly block upper-triangular and nilpotent of index at most LL. This nilpotency guarantees the exact termination of the Neumann series solution after at most LL terms, revealing classical backpropagation to be mathematically equivalent to block back-substitution on an upper bidiagonal system. We formalise F-symmetry -- the condition in which the backward pass perfectly mirrors the forward pass -- identifying orthogonal weight matrices as canonical examples. Through worked numerical examples, we demonstrate how this operator perspective exposes the single-path collapse of strictly feedforward networks and its breakdown in residual architectures. Finally, we leverage this compositional structure to rigorously derive the mechanics of residual networks (gradient highways) and transfer learning (gradient truncation). This framework elevates backpropagation from an algorithmic recipe to a global nilpotent-operator formulation.
Jul 7, 2026

K-ABENA: K-Adaptive Backpropagation with Error-based N-exclusion Algorithm : (Compensated Loss-Based Sample Exclusion with Unbiased Gradient Estimation)

We present K-ABENA (K-Adaptive Backpropagation with Error-based N-exclusion Algorithm), a selective gradient computation framework that reduces per-iteration training cost by excluding a fraction of low-loss ("minor") observations from the backward pass. Its canonical form (v3) combines a defensive-mixture sampling design over the minor set with Horvitz-Thompson inverse-probability reweighting, yielding a design-unbiased Horvitz-Thompson gradient estimator (Lemma 2) and whose self-normalized practical variant carries a bias of order O(1/m) with an explicit constant (Lemma 3). We prove an O(1/sqrt(T)) non-convex convergence guarantee for SGD under the estimator, with an additive term that quantifies the residual bias (Theorem 1). We further prove that uncompensated loss-based selection - a family that includes OHEM, SBP, and the two earlier K-ABENA variants - admits no stationary point at any minimizer where its selection bias is bounded away from zero (Proposition 2), and we quantify this failure empirically: at 0.17% class imbalance, uncompensated variants reach test AUC 0.53-0.62 versus 0.9998 for full-batch SGD, while the compensated estimator attains 0.9991 at identical 28.4% compute savings. On real datasets (Breast Cancer, Digits, Wine, Diabetes) the compensated estimator is statistically indistinguishable from full-batch SGD (paired permutation tests, p >= 0.5; Section 7) while saving 28-54% of per-epoch gradient computation. A biased "regularized mode" (the earlier half-domain variant) is retained as an option with a proven exact bias decomposition (Lemma 5) and quantified contraindications: it collapses to 0.386 accuracy under 40% label noise (baseline: 0.832) and to 0.53 AUC under extreme imbalance. Every advantage and every limitation reported in this paper is either proved or measured; all experiments are CPU-scale (NumPy/scikit-learn) and their scope is stated explicitly.
Jun 22, 2026

FORGE: Fused On-Register Gradient Elimination for Memory-Efficient LLM Training

Reverse-mode differentiation computes every weight gradient, writes it to memory, and only then lets the optimizer read it back. This two-phase schedule sets the memory ceiling of modern training: at the seam between the phases, every layer's gradient is live at once. We argue that this materialized gradient is an artifact of how differentiation is staged, not a quantity that learning requires -- and we eliminate it. FORGE folds the optimizer step into the backward pass and applies it one tile at a time, entirely in registers, so each gradient tile is consumed the instant it is produced and never becomes a tensor. The fusion changes only when the update happens, not what it computes: in full precision the fused step is provably exact -- the identical optimizer update, for every element-wise rule -- and that exactness survives tensor- and sequence-parallel sharding; in the bf16 and 8-bit regimes used in practice it is faithful rather than bit-identical, its deviation bounded and, for the weight store, rendered unbiased by stochastic rounding. Because each gradient tile is born and consumed in the same registers, it is never converted down to bf16 to be stored and read back; FORGE thus preserves the full-precision fidelity that both bf16 and 8-bit optimizers lose to that conversion. Nor is the method tied to one architecture or one optimizer: linear layers are ubiquitous, and FORGE reclaims the gradient memory of any of them under any element-wise rule. Empirically FORGE more than halves the memory of an optimizer step and, at the small batch sizes typical of fine-tuning and continued pretraining, runs about 1.5x faster; integrated into tensor-parallel Megatron-LM it fits 8B training at four times the micro-batch a standard optimizer allows on the same GPUs.
Jun 13, 2026

Automatic Differentiation from Scratch: How PyTorch Computes Gradients in Physics-Informed Neural Networks

This paper traces, with explicit numerical values, how PyTorch's automatic differentiation (AD) engine computes gradients for Physics-Informed Neural Network (PINN) training -- a setting that requires two levels of differentiation: computing the physics derivative y^′(t)=dy^/dt\hat{y}'(t)=d\hat{y}/dt through the network, and computing parameter gradients ∇θL\nabla_θL of a loss that itself depends on y^′(t)\hat{y}'(t). Using a 1-3-3-1 multilayer perceptron and the initial value problem y′(t)+y(t)=0y'(t)+y(t)=0, y(0)=1y(0)=1, we trace the complete pipeline at every node: the computational graph built during the forward pass, the reverse-mode backward traversal that computes all 22 parameter gradients in a single pass, and the graph-on-graph mechanism by which create\_graph=True enables correct differentiation through the physics-informed residual. Every adjoint value is verified against the hand derivations of Tahimi (2026), connecting the P/QP/Q sensitivity framework to the vector--Jacobian products used by PyTorch's autograd engine.
Jun 9, 2026

Overcoming Rank Collapse in Feedback Alignment

Backpropagation (BP) is widely viewed as biologically implausible, in part because it requires feedback weights to be the transpose of forward weights for error propagation. Interestingly, when training a network with fixed random feedback weights to circumvent this issue, learning aligns the forward weights with the feedback weights, leading the backpropagated error signal to become an approximation of the standard gradient used by BP. This process, called Feedback Alignment (FA), occurs in MLPs and very shallow CNNs but does not scale well to deeper architectures. In this work, we first investigated differences between BP and FA models, trained on CIFAR10, specifically focusing on the effective rank of the signal. We found that the FA error has a considerably lower rank and hence is constrained to a lower-dimensional subspace compared to BP, limiting exploration of the parameter space. Motivated by this observation, we evaluated two mechanisms for increasing the effective dimensionality of FA: Muon, an optimiser that orthogonalises weight updates; and hidden activity normalisation, which promotes activation orthogonality. Across larger architectures and benchmarks, we find that these methods consistently improve over FA baselines, for example, on CIFAR100 with a Resnet-18, accuracy increases by 9 percentage points. Our results identify low-dimensional gradient dynamics as a key obstacle to scaling FA and suggest that inducing higher-dimensional update geometry is a promising route toward scaling alternatives to backpropagation.
Jun 8, 2026

Learning Dynamics Reveal a Hierarchy of Weight-Induced Layerwise Gram Metrics

We study feed-forward ReLU networks with fixed readout and quadratic loss, and rewrite gradient descent as a collective dynamics of activation fields and conjugate fields on the training set. Working to first order in the learning rate inside a fixed activation chamber, we derive explicitly the one-, two- and three-hidden-layer cases, and then give the arbitrary-depth recursion. For one hidden layer the activation dynamics closes directly and the residual update is governed by the product of an input Gram matrix and a co-activation/backpropagation Gram matrix. For two hidden layers a conjugate field is required, but no nontrivial pullback Gram metric has yet appeared. For three hidden layers the first weight-induced pullback Gram metric enters the conjugate-field dynamics. At arbitrary depth, activation variations propagate forward through a recursive response operator \cUℓαβ\cU_\ell^{αβ}, while conjugate-field variations propagate backward through an effective transport operator \cMℓαβ\cM_\ell^{αβ}. Their contractions reconstruct a layerwise residual kernel Kαβ(L)=∑ℓ=1LQαβ(ℓ−1)Sαβ(ℓ).K_{αβ}^{(L)}=\sum_{\ell=1}^{L}Q_{αβ}^{(\ell-1)}S_{αβ}^{(\ell)}. The resulting description exposes a duality between push-forward and pullback transport across every layer cut, and identifies the first Gram metrics as the lowest nontrivial terms in a broader hierarchy of activation-conditioned transport operators. We deliberately stop at the level of collective fields, conjugate fields, residual kernels and cut-wise transport metrics, leaving the later tensorial geometric formulation outside the scope of this paper.