Backpropagation-Free Trunk Training via the Split Forward Gradients

Backpropagation (BP) dominates deep learning training, but its reliance on gradients brings inherent troubles -- vanishing and exploding gradients. The pursuit of gradient-free methods has long been a goal in the field of artificial intelligence. This paper shows that indeed the simplest Monte Carlo algorithm implemented on a single GPU -- randomly mutate a parameter, keep it if the loss decreases, otherwise retry -- can practically train deep networks. This gradient-free method does not even need common techniques such as batch normalization or residual connections to directly train sufficiently deep networks. More remarkably, its flexibility extends to several nontrivial scenarios: it enables pure pruning training, supports discrete weights, accommodates unconventional transfer functions such as Gaussian, and reveals the substantial redundancy of deep networks. We have demonstrated its feasibility on deep networks with more than 20 layers, single-hidden-layer wide networks with up to 16,384 hidden neurons, and even a simple Transformer architecture trained on both image classification (MNIST) and character-level language modeling (Tiny Shakespeare). This simple gradient-free method may offer a complementary perspective for understanding the self-organization and learning mechanisms of neural networks, and also provides an alternative route for building physically inspired deep learning systems.

The Forward-Forward algorithm eliminates global gradient flow and full network activations storage. However, in convolutional settings, existing BP-free FF methods significantly under-perform backpropagation on complex benchmarks such as ImageNet-100 and Tiny-ImageNet. We identify this gap as a structural bottleneck in goodness extraction: standard sum-of-squares formulation collapses feature volumes into channel-wise activation energies which omits critical second-order dependencies. To address this, we propose a framework centered on three key components. First, Bi-axis Covariance Goodness(BiCovG) explicitly augments the standard goodness function with structured second-order information along two axes: cross-channel projections that model inter-feature covariance, and nested multi-scale aggregation that encodes spatial correlation statistics. This provides a tractable approximation to covariance-aware goodness without the prohibitive O(C^2) complexity of explicit matrix estimation. Second, a lightweight Logistic Fusion module aggregates layer-wise predictions, amplifying the contribution of deeper representations. Third, the Feature Alignment Layer(FAL) introduces a zero-initialized correction at block boundaries to mitigate representation misalignment in deep locally trained networks. By introducing these three components, we effectively double the depth of viable Forward-Forward learning, extending robust layer utilization from shallow baselines to 16 layer architectures like VGG-16. The resulting BP-free model achieves 73.01% on ImageNet-100 and 50.30% on Tiny-ImageNet. As a practical extension, Hybrid Goodness Blocks control the scope of gradient propagation via configurable block sizes, further narrowing the ImageNet-100 gap to 3.6% and matching BP on Tiny-ImageNet, while still reducing peak memory by approximately 50% relative to BP.

Forward-Forward (FF) training allows each layer to learn from a local goodness criterion. In cumulative-goodness variants, however, later layers can inherit a task that earlier layers have already partially separated. We formalize this phenomenon as layer free-riding: under the softplus FF criterion, the class-discrimination gradient reaching block $d$ decays exponentially with the positive margin accumulated by preceding blocks. We then study three local remedies -- per-block, hardness-gated, and depth-scaled -- that recover current-layer separation measures without relying on backpropagated gradients. On CIFAR-10 and CIFAR-100, these remedies dramatically improve layer-separation statistics, with $4\times$--$45\times$ gains in deeper layers, while changing accuracy by less than one percentage point for non-degenerate training procedures. Tiny ImageNet provides a tougher cross-dataset check for our selected block-wise configuration and reveals the same qualitative gap between layer-health diagnostics and final accuracy. Calibration experiments further show that architecture and augmentation choices have a larger effect on final accuracy than the training-rule modifications studied here. Cumulative free-riding is therefore a real and repairable optimization pathology. Nonetheless, for the FF training rules, architectures, and datasets we study, it is not the dominant factor limiting achievable accuracy.