Local to Global Gradient Equivalence in Neural Networks

50 concepts36 papers1 proposal

Deep learning owes its success to backpropagation, an algorithm that computes global gradients to optimize millions of parameters across deep hierarchies. Yet, from a biological perspective, backpropagation is deeply problematic. It requires non-local information: a neuron would need to know the exact weights of downstream connections (the weight transport problem) and perfectly coordinate distinct, sequential forward and backward passes. For decades, this requirement created a fundamental rift between artificial intelligence and neuroscience.

However, a profound realization has emerged across recent research: the brain does not need to explicitly execute the algorithmic steps of backpropagation to achieve its results. Instead, various biologically plausible architectures exploit local dynamics that are mathematically equivalent to global gradient descent. When these local systems settle into equilibrium, minimize local prediction errors, or balance inhibitory feedback, they implicitly compute the exact global gradients required for deep learning.

Energy Relaxation as a Backward Pass

How can a network propagate error signals backward without a dedicated, artificial backward circuit? The answer lies in the natural relaxation dynamics of physical systems.

In Equilibrium Propagation: Bridging the Gap between Energy-Based Models and Backpropagation, Benjamin Scellier et al. demonstrated that energy-minimizing recurrent networks can compute gradients using a single unified circuit. Instead of a separate backward pass, the network operates in two sequential phases. In the "free phase," the network receives an input and naturally settles into a steady-state equilibrium. In the "nudged phase," a small error signal is applied to the output layer, acting as a physical force that pushes the network into a slightly altered equilibrium.

The core intuition is that the physical displacement of the network's internal states between these two phases contains the exact error information. In Equivalence of Equilibrium Propagation and Recurrent Backpropagation, the authors proved that any system minimizing an energy function through its natural dynamics implicitly executes a form of backpropagation. This equivalence holds strictly in recurrent architectures; Maxence Ernoult et al. showed in Updates of Equilibrium Prop Match Gradients of Backprop Through Time in an RNN with Static Input that these local updates match backpropagation through time (BPTT). By taking the limit of infinitesimal nudging signals, Zhuo Liu et al. confirmed in Toward Practical Equilibrium Propagation: Brain-inspired Recurrent Neural Network with Feedback Regulation and Residual Connections that equilibrium propagation converges precisely to standard backpropagation gradients, establishing a rigorous bridge between physical relaxation and global optimization.

Distributed Inference and Predictive Coding

Another way to bypass a centralized backward pass is to distribute the error computation across the network using local prediction errors. This is the core mechanism of Predictive Coding (PC).

Instead of waiting for an error to propagate from the output layer all the way back to the input, a predictive coding network interleaves state representation and error minimization at every single layer. In [Can the Brain Do Backpropagation?

  • Exact Implementation of Backpropagation in Predictive Coding Networks](/work/can-the-brain-do-backpropagation-exact-implementation), Yuhang Song et al. demonstrated that by utilizing local error nodes to mediate information flow, predictive coding achieves an exact mathematical equivalence to backpropagation.

Why does this local relaxation work globally? As shown in the theoretical framework presented in A Theoretical Framework for Inference Learning, minimizing local prediction errors within a recurrent network effectively optimizes the global loss function through the mechanism of proximal operators. When the network's width is significantly greater than its depth, Francesco Innocenti et al. proved in $\mu$PC: Scaling Predictive Coding to 100+ Layer Networks that these local updates mathematically converge to standard backpropagation gradients, allowing deep brain-inspired architectures to scale efficiently. Similarly, Rasmus Høier et al. showed in Dual Propagation: Accelerating Contrastive Hebbian Learning with Dyadic Neurons that local updates based on internal state differences can collapse two-phase inference into a single step that converges directly to the analytical gradient.

Compartmentalized Cancellation and Ghost Units

At the single-neuron level, the brain solves the challenge of isolating error signals by leveraging the complex geometry of pyramidal cells. Pyramidal neurons possess distinct compartments: basal dendrites receive feedforward data, while apical dendrites receive top-down feedback.

In Dendritic cortical microcircuits approximate the backpropagation algorithm, Joao Sacramento et al. proposed an architecture where dendritic compartments calculate local prediction errors. Crucially, somatostatin-expressing (SST) interneurons learn to predict and cancel out expected top-down feedback. When the feedback matches expectations, the dendritic potential remains quiet. However, when an unexpected global error occurs, the cancellation fails, isolating a pure local error signal that drives synaptic plasticity in exact alignment with backpropagation.

[Top-Down Feedback] ──> [Apical Dendrite] 
                             │  ▲
                             ▼  │  (Inhibitory Cancellation)
                        [SST Interneuron]
                             │
[Feedforward Input]  ──> [Basal Dendrite] ──> [Soma Output]

This concept of feedback cancellation is mathematically powerful. Thomas Mesnard et al. extended this in Ghost Units Yield Biologically Plausible Backprop in Deep Neural Networks, showing that specialized inhibitory "ghost units" can cancel out feedback signals during normal processing, enabling the network to propagate error signals through its own recurrent dynamics. Furthermore, Will Greedy et al. demonstrated in Single-phase deep learning in cortico-cortical networks that by multiplexing forward and backward signals via neural bursting probabilities, local synaptic updates can align perfectly with gradient descent without requiring separate artificial phases. By approximating global objectives like the Evidence Lower Bound via Taylor expansions, Colin Bredenberg et al. showed in Impression learning: Online representation learning with synaptic plasticity that the difference between local dendritic currents and total neuronal firing rates provides a biologically plausible gradient for generative modeling.

Hebbian Plasticity as Implicit Gradient Descent

Even classic, purely local learning rules like Hebbian plasticity and Spike-Timing-Dependent Plasticity (STDP)—which simply correlate pre- and post-synaptic activity—can be viewed as gradient descent on specific global objective functions.

In Spike-timing-dependent Hebbian learning as noisy gradient descent, Niklas Dexheimer et al. mathematically proved that STDP acts as a form of noisy gradient descent on a non-convex loss function, explaining how local spike events guarantee global convergence. This mathematical mapping dates back to early statistical insights; Nathan Intrator demonstrated in A Neural Network for Feature Extraction that biological synaptic plasticity models based on sliding thresholds (BCM theory) are mathematically equivalent to statistical projection pursuit methods designed to find non-linear data structures.

When applied to spiking networks, exact gradient descent naturally aligns with Hebbian correlation. Jane H. Lee et al. showed in Exact Gradient Computation for Spiking Neural Networks via Forward Propagation that exact gradients can be computed during the forward pass, revealing a formal link where weight updates depend strictly on the correlation of activity at firing times. To bridge the gap to deep networks, Roy Abel et al. showed in Biologically Inspired Learning Model for Instructed Vision that a "Counter-Hebbian" learning rule, which correlates activity across reciprocal processing streams, is mathematically equivalent to backpropagation under symmetric weight conditions. Similarly, Alexander Hanuschkin proved in A Hebbian learning rule gives rise to mirror neurons and links them to control theoretic inverse models that local rules incorporating eligibility traces map perfectly to global gradient descent for minimizing motor postdiction errors.

Three-Factor Rules and Temporal Credit Assignment

How does the brain optimize behavior when feedback is delayed or sparse? It modulates local Hebbian plasticity with global reinforcement signals—often called three-factor learning rules.

In Biologically Plausible Multi-dimensional Reinforcement Learning in Neural Networks and Biologically plausible reinforcement learning of continuous actions, Jaldert O. Rombouts et al. demonstrated that combining local pre- and post-synaptic activity with a global reward signal mathematically approximates the gradient of a global objective function. This decentralized optimization can be framed as a collective of independent agents: Stephen Chung proved in MAP Propagation Algorithm: Faster Learning with a Team of Reinforcement Learning Agents that treating every neuron as an RL agent performing maximum a posteriori (MAP) estimation converges to the exact parameter updates of backpropagation.

For temporal tasks, these local updates rely on eligibility traces—biochemical tags that store a memory of recent synaptic activity. Yuhan Helena Liu et al. showed in Can Biologically Plausible Temporal Credit Assignment Rules Match BPTT for Neural Similarity? E-prop as an Example that learning rules utilizing local eligibility traces produce neural representations statistically indistinguishable from those generated by standard backpropagation through time. This temporal alignment is also evident in spiking networks; Johanni Brea et al. demonstrated in Sequence learning with hidden units in spiking neural networks that real-time approximations using low-pass filters achieve performance nearly identical to exact batch-processing methods.

To scale these systems, researchers often integrate global feedback with local rules. Wei Lü highlighted in Training Spiking Neural Networks Using Lessons From Deep Learning how applying BPTT to leaky integrate-and-fire neurons reveals a direct mathematical alignment with biological STDP. This hybrid approach is leveraged by Zihang Shao et al. in EICIL: Joint Excitatory Inhibitory Cycle Iteration Learning for Deep Spiking Neural Networks and Dongcheng Zhao et al. in GLSNN: A Multi-Layer Spiking Neural Network Based on Global Feedback Alignment and Local STDP Plasticity, demonstrating that combining global feedback alignment with local plasticity allows deep spiking networks to match backpropagation performance.

Finally, Mingqing Xiao et al. showed in Training Feedback Spiking Neural Networks by Implicit Differentiation on the Equilibrium State that implicit gradients in equilibrium-based spiking networks align closely with Hebbian correlation principles, bypassing the need for memory-heavy backpropagation through time.

Remarkably, this local-to-global equivalence extends beyond neural tissue: Enrico Facca et al. showed in Physarum-inspired multi-commodity flow dynamics that biological growth dynamics in slime molds adapting edge capacities based on local flow intensity are mathematically equivalent to mirror descent in convex optimization. This underscores a universal principle: biological systems do not need to violate locality to optimize globally; their natural dynamics are already performing the calculus.

Key Foundational Work

Sequence learning with hidden units in spiking neural networks (2011) — contributed 1 concept elements to this direction.
A Neural Network for Feature Extraction (1989) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_biologically_plausible_optimization_5adf32eb,
  title = {biologically plausible optimization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/biologically-plausible-optimization/rdr_5adf32eb70944afc9322115691971151},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_biologically_plausible_optimization_5adf32eb_v1,
  title = {biologically plausible optimization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/biologically-plausible-optimization/rdr_5adf32eb70944afc9322115691971151/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake