In biological systems, time is not measured with the crystalline precision of a quartz watch. Instead, our internal sense of duration is governed by a fundamental psychological principle: the scalar property. This property, a form of Weber’s Law, dictates that the error in our timing estimates grows proportionally with the length of the interval being measured. If you are likely to be off by 100 milliseconds when timing a one-second interval, you will likely be off by a full second when timing a ten-second interval.
Understanding why this linear scaling of uncertainty exists—and how noisy, millisecond-scale neurons produce it—is a central challenge in computational neuroscience.
The Stochastic Accumulator Model
The most intuitive way to think about biological timing is through the lens of stochastic accumulation. Imagine a bucket being filled by several leaky faucets. If the faucets drip at a constant average rate, you can estimate how much time has passed by looking at the water level.
In the brain, this "water" consists of neural spikes. Jonathan Shapiro and John Wearden demonstrated in Reinforcement Learning and Time Perception -- a Model of Animal Experiments that a network of noisy, mutually excitatory neurons naturally acts as an accumulator. Because individual neurons are inherently "jittery," the total spike count (the "water level") isn't perfectly consistent. Crucially, as the interval grows longer, the variance in this cumulative count increases. This stochastic process ensures that timing errors are not a failure of the system, but a direct consequence of its noisy building blocks.
Achieving Scale Invariance
While simple accumulation explains why noise increases over time, it doesn't automatically explain why it increases linearly (the scalar property). To achieve true scale invariance, the system's internal noise must be precisely calibrated.
François Rivest and Yoshua Bengio proposed a solution in Adaptive Drift-Diffusion Process to Learn Time Intervals. They showed that if the variance of the internal noise is proportional to the "speed" of the accumulator (the synaptic weights), the resulting timing precision becomes scale-invariant. This allows an organism to learn "when" an event will occur across vastly different timescales using the same underlying mechanism.
Similarly, Junfeng Zuo and Wenhao Zhang found that multiplicative noise in recurrent neural circuits naturally reproduces Weber’s Law. By modulating the "gain" or intensity of a neural signal, the brain can speed up or slow down its internal clock, with the uncertainty scaling perfectly alongside the speed.
Dedicated Clocks vs. Intrinsic Dynamics
Researchers generally divide timing models into two camps, as noted by Hamit Basgol in Psychological, Computational and Robotic Models of Time Perception:
- Dedicated Models: Specialized structures, like the "late-spiking" neurons in the perirhinal cortex modeled by Katherine Tieu and Thomas H. Brown, act as explicit timers for threat prediction.
- Intrinsic Models: Timing emerges from general neural activity. For instance, Warrick Roseboom and Anil K. Seth suggest that our sense of duration may simply be a byproduct of how many salient changes our perceptual systems process in a given window.
Even artificial agents seem to split this difference. Ben Deverett observed that while recurrent AI agents build internal "clocks," feedforward agents invent "behavioral clocks" by moving in stereotyped patterns to track time.
Rationality in Imprecision
If our internal clocks are so noisy, why haven't we evolved better ones? Fuat Balcı and Patrick Simen argue in Optimal Temporal Risk Assessment that humans and animals are actually "optimal" given their constraints.
Because internal noise increases with duration, the risk of missing a deadline also increases. Organisms compensate for this by becoming more conservative as intervals grow longer. This "irrational" bias is actually a sophisticated form of risk management: by accounting for their own endogenous uncertainty, animals maximize their expected rewards in time-sensitive environments.
Bridging the Timescale Gap
A final mystery is how neurons—which operate on millisecond scales—can time intervals lasting minutes. Patrick Simen proposes a relay race of integrators. By chaining together biological processes with different time constants—from fast neural firing to slow protein synthesis—the brain can bridge the gap between cellular physics and behavioral timing, all while maintaining the robust, scalar-invariant signatures of the biological clock.