In deep learning, global optimization is the undisputed standard. Algorithms like backpropagation rely on a centralized flow of information, where downstream error gradients are meticulously calculated and passed backward to update every weight in the network.
However, physical and biological systems operate under a strict synaptic locality constraint. As Ralph Bourdoukan et al. note, a biological synapse is a microscopic chemical junction with no global awareness; its plasticity is restricted entirely to information available at its immediate connection point, such as pre-synaptic activity and post-synaptic potential. It cannot look across the network to see downstream weights—a limitation known as the "weight transport problem," discussed by Mohamed Akrout et al.—nor can it directly read a global error gradient.
This creates a profound paradox: how can a decentralized network of neurons, each acting on purely local signals, self-organize to solve complex global optimization problems? The answer lies in designing architectures where global objectives naturally decompose into local updates.
Coordinate Transformation via Similarity Matching
To understand how global objectives can be solved locally, consider unsupervised representation learning tasks like Principal Component Analysis (PCA) or Non-Negative Matrix Factorization. The global goal of these algorithms is to reduce dimensionality while preserving the underlying structure or "similarity" of the input data.
Cengiz Pehlevan et al. demonstrated that if you write a global objective function that minimizes the difference between the similarity of the inputs and the similarity of the outputs, a beautiful mathematical decomposition occurs. When you take the gradient of this objective with respect to neural activities and weights, the optimization naturally maps onto the dynamics of a single-layer neural network:
- Feedforward connections learn the principal features of the input data via standard Hebbian plasticity (strengthening connections when pre- and post-synaptic neurons fire together).
- Lateral inhibitory connections develop between the output neurons. To prevent all neurons from converging on the exact same dominant feature, they must compete. The gradient dictates that these lateral connections update via anti-Hebbian rules (strengthening inhibition when neurons fire together).
Because both the feedforward and lateral updates depend exclusively on the concurrent activity of the two neurons a synapse connects, the entire global optimization is executed through strictly local interactions. This is not a loose approximation; Mien Brabeeba Wang proved that local updates like Oja's rule can achieve information-theoretically optimal convergence rates for streaming PCA. By structuring the network with lateral inhibition, the system forces decentralized neurons to collectively solve geometric objectives, such as separating correlated signal sources, as shown by Bariscan Bozkurt et al..
Replacing the Backward Pass with Network Dynamics
Supervised learning in deep networks introduces a harder challenge: credit assignment across multiple layers. How can a synapse in the first layer know how its weight adjustment affects an error at the final output layer without a backward pass?
Instead of building an explicit, non-local backward circuit, we can let the network's own recurrent physics do the math. This is the core intuition behind Equilibrium Propagation, a framework analyzed by Benjamin Scellier et al..
Phase 1: Input (X) ---> [ Hidden Layers Settle ] ---> Prediction (Y)
|
Phase 2: Input (X) ---> [ Hidden Layers Nudged ] <--- Nudge (Y + Error)
In Equilibrium Propagation, the network operates in two temporal phases:
- The Weakly Clamped Phase: The network receives an input and naturally settles into a steady state of neural activity, producing an initial prediction.
- The Nudged Phase: A small error signal is introduced at the output layer as a physical "nudge." This perturbation propagates backward through the existing recurrent connections, altering the steady-state activity of the hidden layers.
The authors proved a remarkable equivalence: the temporal shift in a hidden neuron's activity from the first phase to the second phase matches the exact error gradient calculated by recurrent backpropagation. Consequently, a synapse does not need a global error signal; it only needs to compare the local pre- and post-synaptic activities across these two slices of time.
To prevent feedback signals from corrupting normal forward processing during inference, networks can introduce structural specializations. For example, Thomas Mesnard et al. introduced "Ghost Units"—specialized inhibitory neurons that dynamically cancel out feedback signals during normal inference but allow error perturbations to flow through during learning. By leveraging time and inhibition, global error information is translated into a local, readable format at the single-synapse level.
Modulating the Forward Pass
An alternative way to bypass the backward pass entirely is to embed error information directly into a modified forward pass. The PEPITA framework, proposed by Giorgia Dellaferrera et al., achieves deep credit assignment using two successive forward passes.
In the first pass, the network processes an input normally and computes an error at the output. In the second pass, rather than propagating that error backward, the global error is used to mathematically perturb the original input. The network then runs a second forward pass with this modulated input.
Because the perturbation at the input layer cascades through all subsequent layers, the difference in a neuron's activation between the first and second forward passes reflects its contribution to the final error. Synapses can then execute standard local Hebbian updates based on these two forward states. By shifting the burden of credit assignment from a complex backward architecture to a modulated forward pass, the network satisfies the synaptic locality constraint while maintaining high learning performance.
Ultimately, these diverse mechanisms demonstrate that the synaptic locality constraint is not a barrier to complex learning. Whether through lateral inhibition in similarity matching, temporal phases in equilibrium propagation, or modulated forward passes, global coordination emerges naturally when network structure and physical dynamics are properly aligned with local rules.
Key Foundational Work
Maximising Sensitivity in a Spiking Network (2004) — contributed 1 concept elements to this direction.
Synchrony and Desynchrony in Neural Oscillator Networks (1994) — contributed 1 concept elements to this direction.