Abstract
Human perception remains robust even under challenging conditions where sensory inputs change rapidly over time. A key phenomenon in this process is temporal adaptation—the reduction of neural responses to repeated external stimuli. While widely observed, the functional utility of this "repetition suppression" for recognizing novel inputs has remained unclear. This study combines human behavioral data, electroencephalography (EEG), and deep convolutional neural networks (DCNNs) to demonstrate that temporal adaptation actively improves sensory representations. By adapting to background noise, participants showed faster and more accurate object recognition. These behavioral gains were mirrored by enhanced contrast-dependent EEG responses and improved neural decoding of object information. Computational modeling further identified "divisive normalization" as a biologically plausible mechanism that explains these benefits and provides spatial stability in dynamic environments. Our findings suggest that temporal adaptation is not merely neural fatigue but a functional filter that optimizes the brain for new sensory information.
Introduction
Our visual environment is rarely static or perfectly clear. We often have to identify objects through "noise," such as fog, heavy rain, or cluttered backgrounds. To maintain robust perception, the brain utilizes temporal adaptation, a process where neurons adjust their sensitivity based on recent history. Traditionally, the resulting decrease in neural firing (repetition suppression) was often dismissed as a sign of metabolic fatigue. However, recent theories suggest it may be a functional mechanism designed to optimize the processing of future, novel inputs.
This study investigates whether adapting to a specific background noise pattern "clears the stage" for new objects, effectively boosting their relative saliency. By manipulating the contrast of objects embedded in noise and measuring both human performance and brain activity, the researchers sought to uncover the mathematical computations—whether intrinsic to individual neurons or emergent from network interactions—that enable this robust sensory processing.
Methods
The research team employed a multi-disciplinary approach to link behavior, neural activity, and computation:
- Human Psychophysics: Participants were tasked with categorizing MNIST digits (3, 6, 8, or 9) embedded in pixelated "pink" noise (1/f power spectral density). Before the test image, they viewed an "adapter"—either the same noise pattern, a different noise pattern, or a blank screen. Object contrast was varied to adjust task difficulty.
- Electroencephalography (EEG): 64-channel EEG recorded Event-Related Potentials (ERPs). Multivariate decoding was applied to the EEG data to quantify how much object-specific information could be extracted across different adaptation states.
- Computational Modeling: Task-optimized DCNNs were modified with three candidate adaptation mechanisms:
- Additive Suppression: An intrinsic mechanism where a unit's previous activity is subtracted from its current state.
- Lateral Recurrence: A network-based mechanism where units inhibit their neighbors.
- Divisive Normalization: A multiplicative feedback loop that scales a neuron's response based on its history, a canonical computation in the brain.
Key Results
Behavioral and Neural Enhancement
Human participants were significantly faster and more accurate at recognizing digits when they had adapted to the "same" noise pattern compared to a different one. This behavioral boost was most pronounced as object contrast increased. EEG analysis confirmed that this wasn't just a behavioral fluke; object category could be decoded more accurately and for a longer duration following same-noise adaptation. This suggests that adaptation directly improves the quality and stability of sensory representations in the brain.
Identifying the Neural Computation
When testing the DCNNs, the researchers found that simple intrinsic mechanisms—specifically additive suppression and divisive normalization—best matched human behavioral patterns. These models captured the "interaction" effect where adaptation benefits are tied to object contrast.
Crucially, while additive suppression performed well on static images, divisive normalization was the only mechanism that remained robust when the noise pattern was slightly shifted spatially. This suggests that the multiplicative nature of divisive normalization allows the brain to maintain a stable "template" of the environment even when the input is time-varying and spatially inconsistent.
Discussion
The findings redefine temporal adaptation as a sophisticated signal-processing tool. By suppressing the neural response to predictable background noise, the brain effectively increases the "contrast gain" for new, relevant signals. The study highlights that much of our robust perception may be "baked into" the biophysics of individual neurons through mechanisms like divisive normalization, rather than requiring complex, high-level network loops.
While the EEG data pointed toward parietal involvement in these contrast-dependent effects, the researchers acknowledge the spatial limitations of EEG and suggest that future work using high-resolution fMRI or intracranial recordings could better pinpoint the specific cortical regions involved. Additionally, moving beyond simple digit recognition to complex 3D natural objects will be essential to confirm the universality of these adaptation benefits in real-world scenes.
Summary updated