🧬 biology

Teaching signal synchronization in deep neural networks with prospective neurons

This paper proposes that neurons equipped with adaptive currents can prospectively predict future inputs to synchronize with delayed teaching signals, thereby solving a fundamental timing challenge in hierarchical deep neural networks and enabling efficient learning and memory retention over extended timescales.

Original authors: Nicolas Zucchet, Qianqian Feng, Axel Laborieux, Friedemann Zenke, Walter Senn, João Sacramento

Published 2026-07-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nicolas Zucchet, Qianqian Feng, Axel Laborieux, Friedemann Zenke, Walter Senn, João Sacramento

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine your brain is a massive, high-tech factory trying to assemble a complex toy. The workers (neurons) are great at their jobs, but they are also incredibly slow. They don't just snap their fingers; they have to slowly soak up information like a sponge, taking time to process what they see and hear. This "sponge-like" slowness is actually a superpower: it lets the factory remember things from a few seconds ago, which is essential for working memory.

But here's the glitch: because the workers are so slow, a message from the boss (the "teaching signal" that says "Good job!" or "Try again!") arrives late. By the time the boss's message hits the assembly line, the workers have already moved on to the next step. It's like a coach yelling "Pass the ball!" to a soccer player who has already kicked the ball and is running down the field. The instruction is out of sync, and the team can't learn properly.

For a long time, scientists thought this was just a fundamental limit of how brains work. But this paper suggests a clever workaround: what if the workers could learn to predict the future?

The Crystal Ball Neuron

The authors propose a special kind of neuron that doesn't just react to what's happening now, but uses a bit of "adaptive current" to guess what's coming next. Think of it like a baseball catcher who doesn't wait for the ball to hit their glove to know where it's going; they anticipate the pitch's trajectory and move their glove to where the ball will be.

In the paper's simulations, these "prospective neurons" act like time-travelers. They estimate the future input and adjust their behavior before the signal actually arrives. This allows them to stay perfectly in sync with the boss's instructions, even though the factory is built on slow, leaky sponges.

The "Leaky" Problem vs. The "Prospective" Fix

The paper explicitly argues against the idea that standard, slow neurons can handle this on their own. If you just have a normal "leaky integrator" (a standard neuron that slowly fills up with water), it will always lag behind. The authors ran simulations showing that no matter how you tune it, a standard neuron will always be a step behind the target, creating a "tracking error." It's like trying to chase a moving target while wearing heavy boots; you'll never catch up.

However, when they added this "prospective" ability—essentially giving the neuron a crystal ball—the lag vanished. In their computer models, these adaptive neurons could track the target trajectory perfectly, effectively behaving as if they were instantaneous, even though they were still physically slow.

Does It Actually Work?

The authors didn't just dream this up; they tested it in three different ways, and the results were promising but specific to their simulations:

The Teacher-Student Test: They set up a simple scenario where a "student" network tried to copy a "teacher" network. When the student used standard slow neurons, it failed to learn. But when the student used prospective neurons, it learned just as well as if it were instant. This worked across many different learning rules, not just one specific type.
The Balancing Act: They threw the neurons into a video-game-style challenge: balancing a pole on a moving cart (the famous "Cartpole" task). This is hard because the cart moves fast, and the "reward" (keeping the pole up) comes with a delay. Standard slow neurons failed miserably. But the prospective neurons? They learned to balance the pole for the full 4 seconds of the test, performing just as well as a hypothetical instant brain.
The Memory Mix: The paper also showed that you can mix these two types. You can have a layer of slow, memory-holding neurons (to remember the past) followed by a layer of fast, predictive neurons (to act on the future). This hybrid team successfully solved a task where they had to wait for a "Go" cue before moving, proving that you don't need every neuron to be a time-traveler, just the ones passing the message along.

The Fine Print

It's important to note that while the math looks solid and the simulations are impressive, this is currently a theory backed by computer models. The authors suggest that real biological neurons might have a mechanism similar to this "adaptive current"—perhaps related to how sodium channels inactivate very quickly (in about 10 ms)—but they haven't proven this exists in a living brain yet.

They also found that the system is robust. Even if the "prediction clock" of the neuron is slightly off (mismatched by a small amount), the system still works. But if the prediction is totally wrong, the learning breaks down.

The Bottom Line

This paper suggests that the brain might solve its own "slow-motion" problem not by speeding up, but by getting better at guessing the future. By using a mechanism that acts like a high-pass filter (subtracting the slow, laggy part of the signal), neurons could effectively cancel out their own delays. It's a playful, counter-intuitive idea: to move fast, you don't need to be fast; you just need to know where you're going before you get there.

Technical Summary: Teaching Signal Synchronization in Deep Neural Networks with Prospective Neurons

Problem Statement

The paper addresses a fundamental challenge in learning within hierarchical neural networks: teaching signal synchronization. Working memory requires neurons to slowly integrate inputs over time to maintain information, but these slow dynamics introduce cumulative processing delays across hierarchical layers. Consequently, when teaching signals (such as external rewards or internal error signals) arrive at output-related areas to guide learning, the neural activity in earlier layers has already shifted due to these delays. This creates a temporal mismatch where the teaching signal is out-of-sync with the specific neural activity it is meant to instruct, hindering effective credit assignment.

While standard error backpropagation assumes instantaneous signal transmission, realistic biological models involve leaky integrators with finite time constants ( $\tau$ ). The authors argue that under rapid input evolution, the separation of timescales required for instantaneous approximation fails, causing error signals to arrive too late to effectively update weights in deep networks.

Methodology

Theoretical Framework: Tracking as a Solution

The authors formalize the delay problem as a tracking problem. They define an "ideal" target trajectory ( $s^*_t$ ) representing the neural state that would exist if the system processed inputs instantaneously (i.e., satisfying the equilibrium equation $s^*_t = f_\theta(s^*_t, t)$ ).

Leaky Integrators: Standard leaky dynamics ( $\tau \dot{s}_t = -s_t + f_\theta(s_t, t)$ ) are shown to lag behind the target trajectory. Theorem 1 proves that the tracking error is bounded by $\gamma\tau/\mu$ , where $\gamma$ relates to the input speed and $\mu$ to network geometry. This implies that without strict timescale separation, leaky neurons cannot perfectly track the target, leading to desynchronized learning signals.
Prospective Dynamics: To solve this, the authors propose modifying the dynamics to include a "prospective" input that estimates future states. The proposed dynamics are:
$\tau \dot{s}_t = -s_t + f_\theta(s_t, t) + \tau \frac{d}{dt}f_\theta(s_t, t)$
Theorem 2 demonstrates that this modification allows the state $s_t$ to asymptotically converge to the target trajectory $s^*_t$ , effectively eliminating the tracking error and synchronizing the system with the instantaneous ideal.

Bio-physical Implementation

Recognizing that true differentiators are physically unrealizable, the authors propose a biologically plausible implementation using adaptive neurons.

They model the derivative term $\frac{d}{dt}f_\theta$ using a high-pass filter approximation, realized by subtracting a low-pass filtered version of the input (an adaptation current, $a_t$ ) from the instantaneous input.
The resulting dynamics are:
$\tau \dot{u}_t = -u_t + \left(1 + \frac{\tau}{\tau_a}\right)f_\theta(u_t, t) - \frac{\tau}{\tau_a}a_t$
where $\tau_a$ is the adaptation time constant. This mechanism mimics the effect of prospective dynamics without requiring explicit derivative calculation.
The authors analyze the robustness of this implementation to mismatches between the neuronal time constant ( $\tau$ ) and the adaptation time constant ( $\tau'$ ), showing that performance remains superior to leaky dynamics across a wide range of mismatches.

Learning Experiments

The theoretical properties are validated through three experimental setups:

Teacher-Student Tracking: A feedforward network learns to mimic an instantaneous teacher network. Various learning rules (Backpropagation, Feedback Alignment, Direct Feedback Alignment, Holomorphic Equilibrium Propagation) are tested.
Motor Control (Inverted Pendulum): An agent learns to balance a pole using a continuous-time Advantage Actor-Critic algorithm with purely online learning and no replay buffers.
Memory-Storing Hybrid Networks: A network combining non-prospective leaky neurons (for memory storage) with prospective neurons (for processing) is trained on a delayed reaching task using Real-Time Recurrent Learning (RTRL).

Key Results

Perfect Tracking: Prospective dynamics enable leaky neurons to track the ideal instantaneous trajectory with exponentially converging error, effectively eliminating the delay that plagues standard leaky integrators.
Learning Performance: In the teacher-student setting, prospective dynamics allow online learning algorithms to achieve test losses comparable to instantaneous backpropagation. In contrast, standard leaky dynamics result in significantly higher losses, especially as the input speed increases relative to the neuron's time constant.
Generality: The synchronization benefit holds across diverse learning rules, including those that do not require symmetric weights (Feedback Alignment) and those that rely on oscillatory teaching signals (Holomorphic Equilibrium Propagation).
Robustness: The learning performance remains high even when the adaptation time constant ( $\tau'$ ) mismatches the membrane time constant ( $\tau$ ), provided the mismatch is not extreme.
Motor Control: In the inverted pendulum task, prospective backpropagation enables the agent to learn the task as efficiently as an instantaneous counterpart, whereas reducing the prospective component leads to learning failure.
Memory Integration: In the hybrid network, the combination of leaky neurons (for memory) and prospective neurons (for synchronization) allows the system to solve delayed reaching tasks online. Freezing the memory parameters or removing the prospective component prevents the network from refining movements to match target trajectories.

Significance and Claims

The paper claims to reveal how neural adaptation can solve a critical timing problem in biological learning. By endowing neurons with adaptive currents that enable prospective coding, the brain can potentially synchronize teaching signals with neural activity despite the inherent delays of slow integration and hierarchical processing.

The authors assert that this mechanism:

Provides a theoretical and practical solution for online learning in deep, hierarchical networks where delays are unavoidable.
Offers a biologically plausible implementation of error backpropagation and other learning rules that do not rely on unrealistic instantaneous transmission.
Bridges the gap between the need for working memory (slow integration) and the need for rapid credit assignment (synchronized teaching signals), suggesting that adaptive neurons allow the brain to maintain memory while still learning efficiently from delayed feedback.

The work positions prospective neurons not merely as a theoretical curiosity but as a necessary component for efficient learning in dynamic environments where the timescale of input changes competes with the timescale of neural integration.