Standard deep learning relies on Backpropagation Through Time (BPTT) to train recurrent and temporal networks. While computationally powerful, BPTT suffers from a fundamental flaw when applied to physical systems like neuromorphic hardware or biological brains: it is non-local in both space and time.
To calculate a weight update using BPTT, a system must unroll the network across the entire duration of a sequence, store every intermediate activation state in memory, and then integrate error signals backward from the end of the sequence to the start. As Fernando J. Pineda noted, this requirement for global temporal information makes standard optimization impossible to implement in physical substrates that must process information and learn simultaneously. Furthermore, standard backpropagation requires a neuron to know the exact weights of downstream connections to propagate errors backward—a hurdle known as the "weight transport problem."
To build scalable neuromorphic hardware and understand biological intelligence, we must shift toward spatiotemporal locality. A learning rule is spatiotemporally local if synaptic updates depend exclusively on information available at the specific physical location of the synapse and at the exact time of the update.
Non-Local (BPTT): [Forward Pass: t=0 → t=T] ===(Store History)===≫ [Backward Pass: t=T → t=0]
Spatiotemporal Local: [Inference & Synaptic Update happen concurrently at each interval t]
The Spatial Dimension: Eliminating Global Weight Transport
Spatial non-locality requires a network to coordinate weight information across distant layers. To achieve spatial locality, a learning algorithm must ensure that a synapse's update depends only on the activity of the two units it directly connects.
One compelling way to achieve this is by reframing credit assignment as a physical control problem. In Deep Feedback Control (DFC), introduced by Alexander Meulemans et al., the network does not pass explicit error gradients backward. Instead, a feedback controller applies a continuous driving force that gently nudges the network's activations toward a desired target state. Because the feedback pathways are embedded directly within the network's architecture, the forward weights can optimize themselves using purely local signals to minimize the "effort" or control signal exerted by the controller, as demonstrated by Alexander Meulemans et al.. This eliminates the need for separate, non-overlapping phases for inference and learning.
Alternatively, spatial locality can be achieved through randomized feedback pathways. In Event-Driven Random Back-Propagation (eRBP), developed by Charles Augustine et al., error signals are broadcast through fixed, random feedback weights rather than the exact transpose of the forward weights. When combined with sparse, event-driven communication in Spiking Neural Networks (SNNs), eRBP allows individual synapses to update asynchronously without global coordination.
The Temporal Dimension: Real-Time Online Learning
Temporal non-locality is driven by the memory bottleneck: the need to remember the past to evaluate the present. To achieve temporal locality, a network must perform online learning, updating its parameters in real time as data flows through the system.
1. Exploiting Intrinsic Noise and Dynamics
Instead of tracking deterministic histories, networks can leverage mathematical properties of stochastic systems. Javier R. Movellan demonstrated that in continuous-time diffusion networks (a continuous version of Boltzmann machines), the intrinsic Brownian noise within the network can be exploited to compute exact trajectory gradients using entirely local, instantaneous computations.
2. Architectural Decoupling and Clocks
Another intuition is to separate the representation of time from the representation of content. Amadeus Maes et al. designed an SNN where a recurrent "clock" network discretizes time into sequential neural activity, while a separate read-out layer captures spatial patterns. Because time is explicitly tracked by the architecture itself, complex spatiotemporal sequences (such as bird songs) can be mastered using simple, local synaptic plasticity rules without unrolling history. On a broader scale, Samuel P. Muscinelli showed that a hierarchy of varying neural timescales—leveraging intrinsic properties like spike frequency adaptation—allows networks to learn long, non-Markovian behavioral sequences entirely without external supervision.
3. Linearized Tracking and Predictive Coding
Why does BPTT require storing history? Because when a hidden state influences a future hidden state through a non-linear function, the gradient depends on every step in between. Nicolas Zucchet et al. bypassed this by utilizing independent, linear recurrent modules. Because the recurrence is linear, the influence of past states can be tracked forward in time via a running summary matrix. This allows the network to capture long-range dependencies online with a memory footprint matching a single inference pass.
Furthermore, if a network is trained to predict its own future states, the prediction error itself can serve as a local teaching signal. Mufeng Tang et al. extended predictive coding into the temporal domain, showing that sequential memories can be stored and recalled purely by minimizing local prediction errors at each time step, eliminating the backward pass entirely.
Bridging the Performance Gap
Historically, local learning rules struggled to match the performance of BPTT on complex tasks involving long temporal delays, a limitation highlighted by Yuhan Helena Liu et al.. However, modern formulations close this gap by combining local plasticity with self-supervised objectives.
For example, Yiting Dong et al. introduced PredNext, a framework that forces SNNs to predict future video frames across different augmented views. By structuring the objective around predictive temporal relationships, the underlying spiking layers naturally learn stable, semantically rich representations from raw video data using local rules, matching the performance of fully supervised models. Additionally, incorporating mechanisms like Intrinsic Plasticity (IP)—where individual neurons homeostatically regulate their own excitability—allows local Hebbian learning rules to solve highly non-linear feature discovery problems, as shown by Nicholas J. Butko.
The Physics of Spatiotemporal Locality
Ultimately, spatiotemporal locality is not just a biological constraint; it is a requirement of the physical universe. In standard deep learning, information is assumed to propagate across the network instantaneously.
Alessandro Betti et al. reframed neural learning as an optimal control problem governed by physical Hamiltonian equations. By introducing a finite velocity constant for signal propagation, they proved that learning algorithms naturally become spatiotemporally local: information can only move between immediate neighbors over time. Under this framework, standard backpropagation is revealed to be a physically unrealistic limiting case where the propagation speed is assumed to be infinite.
Infinite Speed (Backprop): Signal reaches all layers instantly.
Finite Speed (Physical): Signal propagates locally: Layer 1 → Layer 2 → Layer 3 over time.
As Anatoli Gorchetchnikov et al. argued, unifying diverse learning phenomena (Hebbian, error-based, and reinforcement learning) into a single mathematical framework governed by strict spatiotemporal constraints is the only viable path for scaling neuromorphic hardware to billions of synapses. By ensuring that updates rely solely on information present at the physical synapse at the exact moment of the event, we can build ultra-low-power, real-time learning systems that operate like the brain.
Key Foundational Work
A Local Algorithm to Learn Trajectories with Stochastic Neural Networks (1993) — contributed 1 concept elements to this direction.
Time Dependent Adaptive Neural Networks (1989) — contributed 1 concept elements to this direction.