Summary:
Researchers in computational neuroscience at the University of Tokyo have built a biologically grounded recurrent spiking neural network that learns to predict, within a single neural population, what event will occur, when it will occur, and how likely it is. Published in Communications Biology, the model abandons biologically implausible global training methods such as backpropagation and instead relies on local learning rules at the readout synapses. The network adapts rapidly to sudden changes in environmental statistics, offering a promising blueprint for neuromorphic hardware and brain-inspired artificial intelligence.
Key Facts:
- Unified predictive architecture: A single population of 1,000 spiking neurons learned to encode simultaneously the identity, timing, and probability of upcoming events, rather than separating these functions into different modules.
- Biologically plausible local learning: The model avoids global backpropagation and broadcast error signals, updating predictions with synapse-local learning at the readouts and reaching reliable performance in fewer than 50 trials per block.
- Fast adaptation to statistical shifts: When event probabilities or temporal delays changed abruptly, the network quickly recalibrated its anticipatory activity, outperforming alternatives based on global least-squares optimization.
Source: University of Tokyo / WPI-IRCN
Everyday perception is predictive: When you spot a yellow light, hear a distant siren, or sense approaching footsteps, your brain forms a compact forecast that includes what will happen, when it will happen, and how likely the event is. Despite how effortlessly humans make such multidimensional predictions, most computational and AI models either split these dimensions into separate subsystems or require learning rules that are hard to reconcile with biological synapses.
A team led by Associate Professor Zenas C. Chao and Academic Specialist Yohei Yamada at the International Research Center for Neurointelligence (WPI-IRCN) developed an alternative: a recurrent spiking neural network (RSNN) that acquires and flexibly maintains a combined representation of identity, timing, and probability using only local, biologically tenable plasticity mechanisms.
Using an error-modulated, attention-gated Hebbian rule applied at readout connections, the network demonstrated that a single recurrent population can learn a complete prediction object and update it online as environmental statistics change. The approach highlights how simple local rules, combined with recurrent dynamics, can reproduce core aspects of anticipatory activity observed in the brain.
Testing the Multi-Event Expectation Task
To probe how a recurrent circuit handles complex expectations, the researchers built an RSNN of 1,000 interconnected spiking neurons and trained it on a Multi-Event Expectation Task. Each trial began with a short sensory cue that signaled one of two possible target events. Each target had a distinct delay window and its own independently adjustable probability. Learning was restricted to the synapses feeding the readout neurons, keeping the recurrent weights unchanged.
Across training blocks of 100 trials, the network developed anticipatory dynamics that encoded all three predictive components without interference:
- Magnitude reflects likelihood: The amplitude of anticipatory firing increased proportionally with the event’s probability, so likely events produced stronger internal signals.
- Temporal shifting: When the expected delay for a target was changed, the timing of the network’s anticipatory burst shifted to align with the new interval.
- Factorized encoding within one circuit: Rather than segregating “what” and “when” into distinct subpopulations, identity and timing organized into overlapping, factorized subspaces within the same neural population.
The network reached most of its predictive accuracy within the first 50 trials of a block, demonstrating rapid, sample-efficient learning that is consistent with adaptive behavior in natural settings.
Dynamic Recovery from Environmental Volatility
Real-world statistics are volatile: delays lengthen, and probable outcomes can become rare. To evaluate flexibility, the team abruptly shuffled timing intervals and outcome probabilities. Predictive error rose immediately after the change but then decayed rapidly as the network updated its internal model.
When compared with global least-squares optimization and models that do not update timing online, the local-learning RSNN showed greater stability and faster recovery. The framework generalized across different numbers of potential events, varying temporal jitter, and alternative teaching-signal shapes, indicating robustness across task conditions.
“We wanted to link a simple everyday intuition about prediction with a learning mechanism that could plausibly operate locally,” said Prof. Chao. “Our results provide a computational foundation for how flexible, multidimensional prediction might arise without globally coordinated error broadcasts.”
Implications for Neuromorphic Hardware and Cognitive Science
The model is presented as a computational proof of concept rather than direct proof of specific physiological mechanisms. Still, it produces concrete, testable hypotheses for experimental neuroscience—particularly concerning mixed-selectivity cortical populations and neuromodulatory systems (for example, dopamine or acetylcholine) that could gate local, error-modulated plasticity.
Beyond basic neuroscience, the architecture has practical appeal for neuromorphic computing and low-power edge AI. By relying on local synaptic plasticity instead of resource-intensive backpropagation through time, the design is well suited to hardware implementations that must learn and adapt to unpredictable real-world data streams in real time while conserving energy.
Funding information
This work was supported by the World Premier International Research Center Initiative (WPI), MEXT, Japan (to Z.C.C.).
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- The journal paper was reviewed in full.
- Additional context was added by staff editors.
About this Neuroscience and AI Research:
- Media Contact: Kazuyo Okada
- Source: University of Tokyo
- Image Credit: Image credited to Neuroscience News
- Original Research (Open Access): Communications Biology (Oct 1, 2026). Title: “Joint encoding of ‘what’ and ‘when’ predictions through error-modulated plasticity in biologically plausible spiking networks.” Authors: Yohei Yamada & Zenas C. Chao.
- DOI: 10.1038/s42003-026-10836-2
Abstract
The brain predicts not only which event will occur, but also when it will occur and with what probability. We call this joint representation of identity, timing, and frequency-based probability a complete prediction object. Existing computational models often treat these dimensions separately or depend on learning rules that lack biological plausibility.
Here, we show that a single population of spiking neurons can acquire and flexibly maintain a complete prediction object through local learning. Using a recurrent spiking network trained with an error-modulated, attention-gated Hebbian rule, we independently varied event identity, latency, and probability. The network developed time-locked anticipatory activity whose amplitude scaled with outcome probability and that rapidly recalibrated when timing or probability statistics changed. Identity and timing self-organized into factorized subspaces within the same neural population.
These findings suggest that mixed-selective cortical populations, together with neuromodulator-gated plasticity, may suffice to jointly encode and continuously update multidimensional predictions within a recurrent circuit.