Estimating and interpreting nonlinear receptive field of sensory neural responses with deep neural network models
Abstract
Our understanding of nonlinear stimulus transformations by neural circuits is hindered by the lack of comprehensive yet interpretable computational modeling frameworks. Here, we propose a data-driven approach based on deep neural networks to directly model arbitrarily nonlinear stimulus-response mappings. Reformulating the exact function of a trained neural network as a collection of stimulus-dependent linear functions enables a locally linear receptive field interpretation of the neural network. Predicting the neural responses recorded invasively from the auditory cortex of neurosurgical patients as they listened to speech, this approach significantly improves the prediction accuracy of auditory cortical responses, particularly in nonprimary areas. Moreover, interpreting the functions learned by neural networks uncovered three distinct types of nonlinear transformations of speech that varied considerably from primary to nonprimary auditory regions. The ability of this framework to capture arbitrary stimulus-response mappings while maintaining model interpretability leads to a better understanding of cortical processing of sensory signals.
Data availability
Source data files have been provided for Figures 1-3. Raw data cannot be shared as we do not have ethical approval to share this data. To request access to the data, please contact the corresponding author.
Article and author information
Author details
Funding
National Institutes of Health (NIDCD-DC014279)
- Menoua Keshishian
- Hassan Akbari
- Bahar Khalighinejad
National Institute of Mental Health
- Jose L Herrero
- Ashesh D Mehta
The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.
Ethics
Human subjects: All research protocols were approved and monitored by the institutional review board at the Feinstein Institute for Medical Research (IRB-AAAD5482), and informed written consent to participate in research studies was obtained from each patient before electrode implantation.
Copyright
© 2020, Keshishian et al.
This article is distributed under the terms of the Creative Commons Attribution License permitting unrestricted use and redistribution provided that the original author and source are credited.
Metrics
-
- 4,044
- views
-
- 481
- downloads
-
- 57
- citations
Views, downloads and citations are aggregated across all versions of this paper published by eLife.
Download links
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Further reading
-
- Neuroscience
Our movements result in predictable sensory feedback that is often multimodal. Based on deviations between predictions and actual sensory input, primary sensory areas of cortex have been shown to compute sensorimotor prediction errors. How prediction errors in one sensory modality influence the computation of prediction errors in another modality is still unclear. To investigate multimodal prediction errors in mouse auditory cortex, we used a virtual environment to experimentally couple running to both self-generated auditory and visual feedback. Using two-photon microscopy, we first characterized responses of layer 2/3 (L2/3) neurons to sounds, visual stimuli, and running onsets and found responses to all three stimuli. Probing responses evoked by audiomotor (AM) mismatches, we found that they closely resemble visuomotor (VM) mismatch responses in visual cortex (V1). Finally, testing for cross modal influence on AM mismatch responses by coupling both sound amplitude and visual flow speed to the speed of running, we found that AM mismatch responses were amplified when paired with concurrent VM mismatches. Our results demonstrate that multimodal and non-hierarchical interactions shape prediction error responses in cortical L2/3.
-
- Neuroscience
Recent studies suggest that calcitonin gene-related peptide (CGRP) neurons in the parabrachial nucleus (PBN) represent aversive information and signal a general alarm to the forebrain. If CGRP neurons serve as a true general alarm, their activation would modulate both passive nad active defensive behaviors depending on the magnitude and context of the threat. However, most prior research has focused on the role of CGRP neurons in passive freezing responses, with limited exploration of their involvement in active defensive behaviors. To address this, we examined the role of CGRP neurons in active defensive behavior using a predator-like robot programmed to chase mice. Our electrophysiological results revealed that CGRP neurons encode the intensity of aversive stimuli through variations in firing durations and amplitudes. Optogenetic activation of CGRP neurons during robot chasing elevated flight responses in both conditioning and retention tests, presumably by amplifying the perception of the threat as more imminent and dangerous. In contrast, animals with inactivated CGRP neurons exhibited reduced flight responses, even when the robot was programmed to appear highly threatening during conditioning. These findings expand the understanding of CGRP neurons in the PBN as a critical alarm system, capable of dynamically regulating active defensive behaviors by amplifying threat perception, and ensuring adaptive responses to varying levels of danger.