Abstract
Tasks that require information about the world imply a trade-off between the time spent on observation and the variance of the response. In particular, fast decisions need to rely on uncertain information. However, standard estimates of information processing capabilities, such as the dynamic range, are defined based on mean values that assume infinite observation times. Here, we show that limiting the observation time results in distributions of responses whose variance increases with the temporal correlations in a system and, importantly, affects a system's confidence in distinguishing inputs and thereby making decisions. To quantify the ability to distinguish features of an input, we propose several measures and demonstrate them on the prime example of a recurrent neural network that represents an input rate by a response firing averaged over a finite observation time. We show analytically and in simulations that the optimal tuning of the network depends on the available observation time, implying that tasks require a "useful" rather than maximal sensitivity. Interestingly, this shifts the optimal dynamic regime from critical to subcritical for finite observation times and highlights the importance of incorporating the finite observation times concept in future studies of information processing capabilities in a principled manner.
Introduction
The criticality hypothesis in neuroscience suggests that biological systems are most efficient when tuned to a "sweet spot" between order and chaos—the critical point (where the control parameter $\lambda = 1$). At this point, systems are theoretically thought to maximize their sensitivity and dynamic range. However, traditional mathematical proofs for this optimality rely on the assumption of "infinite observation time," where responses are averaged over an eternity to eliminate noise.
In biological reality, animals must make split-second decisions. When observation time is limited, internal fluctuations and "critical slowing down" become dominant. This paper addresses the gap between theoretical infinite-time optimality and the practical constraints of finite-time decision-making, investigating how finite observation time ($T$) shifts the optimal dynamical regime of a system from critical to subcritical.
Methods and The Model
The authors modeled a recurrent neural network (RNN) consisting of probabilistic integrate-and-fire neurons. The network's dynamical state is controlled by a parameter $\lambda$, representing the largest eigenvalue of the connectivity matrix.
The network receives an input rate ($h$) and produces an output firing rate ($o_T$) averaged over a finite time window ($T$). To evaluate performance under these constraints, the authors introduced "variability-aware" metrics:
- $\epsilon$-discriminable inputs ($n_d$): The number of distinct input levels a system can reliably distinguish given a specific error threshold ($\epsilon$).
- Processing Dynamic Range ($\Delta$): A measure of the input range that can be distinguished from noise and saturation within a finite time.
The study utilized large-scale simulations ($N=10,000$ neurons) and analytical Fokker-Planck equations to predict how fluctuations "smear" signals at different levels of criticality.
Key Results: The Sensitivity-Confidence Trade-off
The core finding is a fundamental trade-off between sensitivity and confidence. Near the critical point, the system is highly sensitive—small input changes lead to large changes in mean output. However, criticality also induces long-lasting fluctuations. In a short observation window, these fluctuations make the output unreliable, causing the system to lose "confidence" in its signal.
As shown in the results, for any finite $T$, the optimal state for discriminating inputs shifts away from the critical point into the subcritical regime ($\lambda < 1$). In this state, the system is slightly less sensitive but significantly more stable, resulting in tighter output distributions that allow for better input discrimination. As the observation time $T$ increases, the optimal $\lambda$ gradually approaches the critical point, but for typical biological timescales (e.g., $T = 1$s), the optimal regime remains subcritical.
Discussion and Biological Relevance
The authors argue that neural systems do not require maximal sensitivity, but rather useful sensitivity. If high sensitivity comes at the cost of high noise, it becomes counterproductive for decision-making. This suggests that the brain may purposefully avoid true criticality to remain reliable in fast-paced environments.
Future work may explore how high-dimensional population coding or more complex network architectures (including inhibitory neurons) might mitigate the noise issues found at criticality. Additionally, different tasks, such as those requiring long-term memory, might still benefit from the long-range correlations found closer to the critical point, suggesting that the "optimal" state is task-dependent.
Summary updated