Incorporating neuronal fatigue in deep neural networks captures dynamics of adaptation in neurophysiology and perception

Kasper Vinken, Xavier Boix, Gabriel KreimanbioRxiv 2019 · DOI

TL;DR

This paper proposes that neuronal "fatigue" – a biophysically plausible intrinsic property of neurons – can explain key adaptation phenomena in visual processing, from neural responses to perception. By incorporating this fatigue mechanism into a deep neural network, they demonstrate it can replicate complex adaptation effects like repetition suppression and perceptual aftereffects, and even improve object recognition.

What Problem Does This Paper Address?

Adaptation is a fundamental aspect of how our visual system processes information, allowing it to adjust to changing environments and focus on novel stimuli. However, the underlying mechanisms of adaptation remain a puzzle. Does it arise from complex circuit-level interactions, or is it a simpler, intrinsic property of individual neurons? This paper tackles this question by investigating whether a neuronally intrinsic, biophysically plausible "fatigue" mechanism is sufficient to explain the diverse manifestations of visual adaptation observed in both neurophysiology and perception.

Methods

The core of the study involves building and testing a deep convolutional neural network (CNN) model inspired by the AlexNet architecture. The key innovation is the implementation of an "exponentially decaying fatigue mechanism" within each unit of the network. This mechanism models how a neuron's responsiveness decreases with sustained input and recovers when the input is removed. The authors then compare the model's outputs to neurophysiological recordings from animal studies and behavioral data from human psychophysics experiments. They also explore training the network's adaptation parameters to optimize recognition performance and compare their fatigue model to a recurrent neural network approach.

Datasets

  • ImageNet dataset: Used for pre-training the AlexNet architecture.
  • Macaque IT cortex recordings: Data from previous studies on face stimuli.
  • Rat visual cortex recordings: Data from previous studies on oddball sequences.
  • Doodle dataset: A custom dataset of hand-drawn doodles used for psychophysics experiments and model fine-tuning.
  • Webmorph: Used to generate face stimuli for the face-gender aftereffect experiment.

Key Insights and Ideas

  • Neuronal fatigue as a unifying mechanism: The paper strongly suggests that a simple, intrinsic neuronal fatigue mechanism can account for a wide range of adaptation phenomena, bridging the gap between cellular biophysics and network-level perception.
  • Beyond circuit-level explanations: While recurrent networks can model adaptation, the authors demonstrate that intrinsic fatigue offers a more robust and generalizable solution, particularly when faced with variations in input statistics.
  • Adaptation enhances efficiency and novelty detection: The model shows that adaptation, by reducing responses to predictable stimuli, can improve object recognition and implicitly highlight novel or unexpected inputs.
  • Cascading adaptation: The study provides evidence that adaptation effects propagate and accumulate across layers in a hierarchical network, mirroring observations in the visual hierarchy.

Key Results

  • Repetition Suppression: The fatigue model successfully replicated stimulus-specific repetition suppression in neural responses, matching data from macaque IT cortex.
  • Perceptual Aftereffects: The model demonstrated classic perceptual aftereffects, such as the tilt aftereffect and face-gender aftereffect, by showing shifts in decision boundaries after adaptation.
  • Improved Object Recognition: Adapting to a prevailing noise pattern significantly improved object recognition performance in both human participants and the computational model.
  • Robustness: The intrinsic fatigue mechanism generalized better to different noise patterns compared to a recurrent neural network solution, highlighting its robustness.
  • Parameter Optimization: Adaptation parameters (α and β) could be trained to maximize recognition performance, leading to high accuracy (97.9%) in the doodle classification task.

Limitations and Future Work

  • Single Time Scale: The current model uses an exponential decay for fatigue, limiting it to a single time scale. Real-world adaptation operates across a range of time scales, and future work could explore more complex models (e.g., power-law adaptation).
  • Neuron-Level Implementation: Adaptation is implemented at the whole-neuron level. Future research could investigate synapse-specific adaptation mechanisms.
  • Suppressive Effects Only: The study focused on suppressive effects of adaptation. Future work could explore response facilitation and how it interacts with fatigue.
  • Circuit-Level Interactions: While the model demonstrates the sufficiency of intrinsic fatigue, it did not capture response enhancements observed in some neurophysiological data. Future iterations could incorporate circuit-level computations to explore these interactions.

Summary updated

Abstract

Abstract Adaptation is a fundamental property of the visual system that molds how an object is processed and perceived in its temporal context. It is unknown whether adaptation requires a circuit level implementation or whether it emerges from neuronally intrinsic biophysical processes. Here we combined neurophysiological recordings, psychophysics, and deep convolutional neural network computational models to test the hypothesis that a neuronally intrinsic, biophysically plausible, fatigue mechanism is sufficient to account for the hallmark properties of adaptation. The proposed model captured neural signatures of adaptation including repetition suppression and novelty detection. At the behavioral level, the proposed model was consistent with perceptual aftereffects. Furthermore, adapting to prevailing but irrelevant inputs improves object recognition and the adaptation computations can be trained in a network trained to maximize recognition performance. These results show that an intrinsic fatigue mechanism can account for key neurophysiological and perceptual properties and enhance visual processing by incorporating temporal context.
BibTeX
@misc{vinken2019incorporating,
  author = {Kasper Vinken and Xavier Boix and Gabriel Kreiman},
  title = {{Incorporating neuronal fatigue in deep neural networks captures dynamics of adaptation in neurophysiology and perception}},
  year = {2019},
  doi = {10.1101/642777},
  howpublished = {bioRxiv},
}
Figure