Summary:
Computational neuroscientists at the University of Tokyo have developed a biologically plausible recurrent spiking neural network capable of simultaneously predicting an event’s identity, precise timing, and probability within a single neuron population.
Published in Communications Biology, the model bypasses biologically unrealistic global training methods like backpropagation, relying instead on local learning rules at readout connections. The network adapted dynamically to sudden environmental shifts, providing a new blueprint for neuromorphic computing and brain-inspired artificial intelligence.
Key Facts:
- Unified Predictive Architecture: A single population of 1,000 spiking neurons learned to concurrently encode “what” an event will be, “when” it will occur, and “how likely” it is, rather than segregating these tasks into isolated computational modules.
- Biologically Plausible Local Learning: The model avoids global backpropagation or broadcast error signals, using local synaptic learning rules at the readout layer to update predictions within 50 trials.
- Rapid Adaptation to Statistical Shifts: When event probabilities or temporal delays were suddenly altered, the network quickly re-stabilized its anticipatory representations, outperforming traditional global least-squares approaches.
Source: University of Tokyo / WPI-IRCN
Everyday perception is fundamentally anticipatory. When a driver spots a yellow traffic light, an approaching siren sounds in the distance, or footsteps echo in a hallway, the human brain instantly constructs a multifaceted forecast. It predicts what sensory event is coming, exactly when it will occur, and the statistical likelihood of its arrival.
Despite how effortlessly the brain makes these calculations, computational neuroscience and artificial intelligence have struggled to replicate them biologically. Traditional machine learning models typically segregate these dimensions into separate, specialized modules or rely on backpropagation—a mathematically powerful training algorithm that relies on globally coordinated feedback signals difficult to reconcile with the physical wiring and metabolic constraints of biological synapses.
Now, a research team led by Associate Professor Zenas C. Chao alongside Academic Specialist Yohei Yamada at the International Research Center for Neurointelligence (WPI-IRCN) at the University of Tokyo has introduced a unified, biologically grounded alternative.
Published in Communications Biology, the researchers demonstrated that a single recurrent population of spiking neurons can simultaneously learn and update event identity, timing, and probability using strictly local learning rules.
“Prediction in everyday life is inherently multidimensional,” said Prof. Chao. “Our results show computationally that a single recurrent spiking population can learn what is expected, when it is expected, and how likely it is, while updating these predictions when environmental statistics change.”
Testing the Multi-Event Expectation Task
To examine how neural circuits manage complex expectations, the Tokyo team constructed a recurrent spiking neural network (RSNN) consisting of 1,000 interconnected neurons.
They designed a “Multi-Event Expectation Task” where the network was presented with a brief sensory cue that predicted one of two subsequent target events. Crucially, each event had its own discrete delay window, and the probability of each event occurring was independently modulated. Learning was confined strictly to the synaptic weights of the network’s readout connections.
When trained across 100-trial blocks, the network integrated all three predictive dimensions without interference:
- Magnitude Reflects Likelihood: When the probability of an event increased, the internal firing activity representing that expectation scaled up proportionally.
- Temporal Shifting: When the anticipated delay changed, the network dynamically adjusted the timing of its anticipatory burst to match the new interval.
- Factorized Within One Circuit: Instead of assigning “what” and “when” to separate clusters of cells, the network formed factorized, overlapping patterns within the exact same neural population.
Remarkably, the network achieved the bulk of its predictive precision within the first 50 trials of a training block, illustrating rapid, sample-efficient learning.
Dynamic Recovery from Environmental Volatility
Real-world environments are constantly shifting: a delay might lengthen or a likely outcome can suddenly become rare.
To test flexibility, the investigators abruptly scrambled the timing intervals and outcome probabilities. Prediction errors spiked immediately following the change, but the network rapidly adapted its internal model.
When benchmarked against alternative computational frameworks—including global least-squares optimization and models lacking online temporal updates—the local learning network displayed superior stability and faster recovery times. The framework also generalized across varying numbers of potential events, temporal jitters, and teaching-signal shapes.
“We wanted to connect a simple everyday idea about prediction with a learning mechanism that could operate locally,” Prof. Chao explained. “The model provides a computational foundation for investigating how flexible prediction might arise without globally coordinated learning.”
Blueprint for Neuromorphic Hardware and Cognitive Science
While the authors emphasize that the model is a computational proof of concept rather than direct physiological proof of how biological brains operate, it offers concrete, testable hypotheses for neuroscientists investigating cortical circuits and neuromodulatory systems (such as dopamine or acetylcholine) that convey local prediction errors.
Beyond basic biology, the discovery holds significant promise for neuromorphic computing and low-power edge AI hardware. Because the architecture relies on local synaptic plasticity rather than resource-heavy backpropagation through time, it offers a hardware-friendly design for autonomous chips that must learn, anticipate, and adapt to unpredictable real-world data streams in real time.
Funding information
This work was supported by World Premier International Research Center Initiative (WPI), MEXT, Japan (to Z.C.C.).
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- Journal paper reviewed in full.
- Additional context added by our staff.
About this Neuroscience and AI Research:
- Media Contact: Kazuyo Okada
- Source: University of Tokyo
- Image Credit: Image credited to Neuroscience News
- Original Research is Open Access: Communications Biology (Oct 1, 2026). “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 what will occur, but also when it will occur and with what probability. We refer to this joint representation of identity, timing, and frequency-based probability as a complete prediction object. Existing computational models typically treat these dimensions separately or rely on biologically implausible learning rules.
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 manipulated event identity, latency, and probability. The network developed time-locked anticipatory activity whose amplitude scaled with outcome probability and rapidly recalibrated when timing or probability statistics changed. Identity and timing self-organized into factorized subspaces within a shared neural population.
These results suggest that mixed-selective cortical populations, coupled with neuromodulator-gated plasticity, may be sufficient to jointly encode and update multidimensional predictions within a recurrent circuit.
