Kickback cuts Backprops red-tape: Biologically plausible credit assignment in neural networks

Error backpropagation is an extremely effective algorithm for assigning credit in artificial neural networks. However, weight updates under Backprop depend on lengthy recursive computations and require separate output and error messages – features not shared by biological neurons, that are perhaps unnecessary. In this paper, we revisit Backprop and the credit assignment problem. We first decompose Backprop into a collection of interacting learning algorithms; provide regret bounds on the performance of these sub-algorithms; and factorize Backprop’s error signals. Using these results, we derive a new credit assignment algorithm for nonparametric regression, Kickback, that is significantly simpler than Backprop. Finally, we provide a sufficient condition for Kickback to follow error gradients, and show that Kickback matches Backprop’s performance on real-world regression benchmarks.

💡 Research Summary

The paper tackles the long‑standing criticism that standard error backpropagation (Backprop) is biologically implausible because neurons would need to emit two distinct signals—forward activations and backward error messages—while real cortical neurons appear to use a single spiking signal. The authors begin by mathematically decomposing Backprop into a collection of local learning agents. Each hidden unit is modeled as a rectilinear (positive or negative ReLU) neuron that incurs a “rectilinear loss” defined as the product of an externally supplied scalar feedback ϕ and the unit’s activation. Theorem 1 shows that the weight update prescribed by Backprop for a hidden unit is exactly the gradient descent step on this local loss, establishing that Backprop is essentially a set of coordinated local optimizers glued together by recursively computed error signals.

Next, Theorem 2 provides a regret bound for these local learners that holds for any sequence of inputs and scalar feedback, even under adversarial conditions. The bound guarantees that the average loss incurred on the time steps when a unit fires converges to the loss of the best fixed weight vector in hindsight, at a rate O(1/√|F|). This result is stronger than typical i.i.d. analyses because, in deep networks, the distribution of inputs to deeper layers is not independent of the learning dynamics.

The core technical insight arrives in Theorem 3, where the authors prove that when a network has a single scalar output (the setting of non‑parametric regression), the error signal δ_j received by any hidden unit j factorizes into the product of the global error β = ∂E/∂x_o and the total influence π_j, which is the sum over all paths from j to the output of the product of weights along each path. Computing π_j is costly and biologically unrealistic.

Motivated by this factorization, the authors introduce Kickback, a new credit‑assignment algorithm that truncates the error signal: instead of using the full total influence π_j, Kickback uses only the immediate influence τ_j = Σ_k w_jk·1_k, i.e., the weighted sum of connections from j to the next layer, multiplied by the same global error β. The resulting update rule is ∆w_ij ∝ –β·τ_j·x_i·1_j. Thus, each neuron needs only its own activation, the global scalar error, and a simple local statistic about its outgoing weights; no recursive back‑propagation of errors is required.

Because Kickback no longer follows the exact gradient of the network loss, the authors identify a sufficient condition called “coherence”: a node is coherent if τ_j > 0, meaning its outgoing influence is positive. If every node in the network is coherent, Theorem 4 proves that, with a sufficiently small learning rate, the sign of the Kickback feedback matches that of the true Backprop feedback, guaranteeing that each update reduces the overall loss. Coherence can be enforced by constraining the sign of weights according to the sign of the target neuron (positive weights to positive‑sign neurons, negative to negative).

The paper also draws a direct line to neurobiology via the “selectron” model, which emerges from taking the fast‑time‑constant limit of the Spike‑Response Model (SRM) combined with neuromodulated Spike‑Timing‑Dependent Plasticity (STDP). In the selectron, a global neuromodulatory signal ν (analogous to dopamine) and pre‑synaptic activity x_i drive weight updates of the form ∆w_ij ∝ ν·x_i·1_j, which is mathematically identical to gradient ascent on a rectilinear reward function. By setting ϕ = –ν, the selectron’s reward aligns with the rectilinear loss used in the theoretical analysis. Consequently, the three components of Kickback—global error β, immediate influence τ_j, and pre‑synaptic activity x_i—map onto known biological substrates: β to reward‑prediction error signals (dopamine), τ_j to NMDA‑mediated back‑connections that modulate plasticity, and x_i to the presynaptic spike train. The indicator 1_j ensures that only active (spiking) neurons undergo synaptic change, mirroring the post‑synaptic gating in STDP.

Empirically, the authors evaluate Kickback on several standard regression benchmarks (Boston Housing, Concrete Strength, Energy Efficiency, Yacht Hydrodynamics) using multilayer perceptrons with two and three hidden layers. They compare training curves, final mean‑squared error, and convergence speed against conventional Backprop. Results show virtually identical performance: Kickback reaches the same error levels, converges in a comparable number of epochs, and exhibits smoother error‑signal propagation because it avoids the costly recursive computation of δ_j. Moreover, the computational overhead of Kickback is substantially lower, as each layer only needs to broadcast its τ_j once per forward pass.

In summary, the paper makes four major contributions: (1) a novel decomposition of Backprop into local rectilinear learners, (2) a regret bound that validates these learners under arbitrary feedback, (3) the Kickback algorithm that simplifies error signaling while preserving performance under a coherence condition, and (4) a biologically grounded interpretation linking Kickback’s components to neuromodulatory and synaptic mechanisms. By providing both rigorous theoretical guarantees and practical experimental evidence, the work bridges a gap between deep learning theory and neurobiological plausibility, suggesting a pathway toward more brain‑like learning algorithms without sacrificing the efficiency of modern gradient‑based methods.