When humans and animals estimate the passage of time, their errors are not constant. Instead, the uncertainty of a time estimate scales proportionally with the duration of the interval being measured—a phenomenon known in psychology as the scalar property of interval timing, or Weber’s Law. If you tend to miss a 5-second timer by half a second, you will likely miss a 50-second timer by 5 seconds.
While artificial intelligence systems often maintain perfectly constant error rates or unrealistic precision across timescales, biological networks are constrained by noise. A major focus in computational neuroscience is understanding how this scalar variability emerges naturally from the stochastic dynamics of neural populations.
Accumulating Noisy Spikes
The most intuitive way to build a biological clock is through an accumulator model. Imagine a leaky bucket filling with water droplets; by measuring the volume of water, you can estimate how much time has elapsed. In the brain, this can be implemented by a network of mutually excitatory, noisy neurons whose cumulative spike count serves as the internal metric of time.
As demonstrated by Shapiro et al. (2001), this stochastic accumulation process inherently links duration to variability. Because individual neural spikes occur stochastically, counting spikes over longer intervals means accumulating more variance. The authors showed that combining this noisy accumulator with standard reinforcement learning algorithms perfectly replicates biological timing behavior, suggesting that timing errors are a direct consequence of inherent cellular noise rather than a failure of the system.
However, individual neurons have brief intrinsic time constants, making it difficult to accurately track long intervals (seconds to minutes) without the system saturating or decaying. To solve this, Simen et al. (2011) proposed that the brain utilizes a "relay race" of sequential integrators. By linking multiple integration stages that span different biological timescales—from fast neural firing to slow protein synthesis—the network can achieve robust, scale-invariant timing without requiring biologically implausible precision from any single component.
The Mathematics of Noise Propagation
To understand exactly why timing variance scales with duration, we must look at how noise propagates through feedforward neural sequences, often modeled as "synfire chains." When an activity wave propagates through a chain of neurons, how do local fluctuations accumulate?
Obeid et al. (2020) provided an analytical framework to decompose this variability into two distinct scaling regimes:
- Local Noise (Linear Variance Scaling): If each stage in a feedforward sequence suffers from independent, local stochasticity, the timing error accumulates like a classic random walk. Here, the timing variance scales linearly with the duration ($\sigma^2 \propto T$), meaning the standard deviation scales with the square root of time ($\sigma \propto \sqrt{T}$).
- Global Noise (Quadratic Variance Scaling): If the system experiences trial-to-trial fluctuations in global parameters—such as systemic neural excitability or synaptic fatigue—the speed of the entire sequence shifts uniformly. This causes the timing error itself to scale linearly with duration, resulting in a quadratic increase in variance ($\sigma^2 \propto T^2$).
Distinguishing between these scaling laws is crucial for identifying the underlying noise processes in biological circuits. When behavioral timing errors occur, they are frequently driven by these macro-scale scaling inaccuracies—such as a premature termination of the sequence—rather than a complete breakdown of the sequential pattern itself, a principle observed in prefrontal cortex dynamics by Li et al. (2024).
Recurrent Dynamics and Multiplicative Noise
While feedforward chains offer a clear view of sequence propagation, the brain frequently relies on recurrent neural networks (RNNs) to generate flexible behavior. How do recurrent circuits maintain a strict proportional relationship between time and error?
The answer lies in how control signals interact with noise. Zuo et al. (2023) demonstrated that incorporating multiplicative noise into gain-modulated recurrent circuits naturally reproduces Weber's Law. In these networks, a central control signal adjusts the "gain" (or intensity) of neural activity to speed up or slow down a behavioral sequence. If the noise in the circuit is multiplicative—meaning it scales up as the control signal scales up—the timing variability remains perfectly proportional to the duration of the interval.
This connection between architectural constraints and biological timing is so robust that it emerges spontaneously in artificial systems. Deverett et al. (2019) found that when deep reinforcement learning agents are trained to track intervals, recurrent agents naturally develop internal "clocks" that mimic primate neural scaling. This suggests that scalar variability is not a biological idiosyncrasy, but a fundamental property of recurrent information processing under resource and noise constraints.
Time as an Emergent Statistical Property
The models discussed so far assume the existence of a dedicated clock or sequence generator. However, a compelling alternative perspective argues that the brain does not need a specialized timing circuit at all.
Because biological systems are constantly flooded with noisy, temporally structured sensory and internal signals, elapsed time can be treated as a statistical inference problem. As argued by Ahrens et al. (2007), temporal information is implicitly encoded within the statistical structure of ongoing, non-dedicated neural processes. Downstream networks can infer how much time has passed simply by reading out how the statistical properties of these ubiquitous signals evolve over time.
As reviewed by Haß (2016), leveraging the inherent noise of distributed neural activity naturally accounts for the scalar property observed in behavior. By framing time perception as the statistical decoding of ongoing noise, these models elegantly explain why our sense of time stretches and deforms alongside the fluctuating state of the brain.
Key Foundational Work
Reinforcement Learning and Time Perception -- a Model of Animal Experiments (2001) — contributed 2 concept elements to this direction.
Perirhinal-amygdala circuit-level computational model of temporal encoding in fear conditioning (1999) — contributed 1 concept elements to this direction.