Stephen V. David

Academic peer review is fundamental for scientific knowledge dissemination, and various initiatives are exploring how the peer review process could be more open, efficient and rewarding. We report an exploratory study where a live community-based review approach was integrated into the editorial workflow of an academic journal (Current Research in Neurobiology; CRNEUR). This study was conducted with five manuscripts ('cases') submitted as preprints, which underwent Live Review-a structured collaborative review session led by PREreview, an organization with the mission to advance equity and openness in scholarly evaluation. With each case, PREreview team members facilitated a 90-min online discussion where registered participants provided real-time discussion and worked together on the online structured peer review document. Authors could join as observers or to answer questions. Participants then volunteered to write up the session notes into a final review and summary statement. Review participants had the option to sign the review. The finalized review was then published on PREreview's open preprint review platform approximately two weeks after the Live Review session, and it was assigned a CC BY 4.0 license and Digital Object Identifier (DOI) linked to the DOI of the reviewed preprint allowing reviewers to be recognized for their contribution. The published review was then incorporated into CRNEUR's editorial workflow to inform editorial decisions and manuscript outcomes. We quantified the speed to first and final editorial decision of the community review (n = 5) in comparison to a larger sample (n = 27) of articles that went through a standard review process at CRNEUR during the same timeframe. First decision times in days after manuscript submission of the Live Reviews were within the Inter Quartile Range (IQR) of the standard review process (community review: median = 75, IQR: 41.3; standard review: median 92, IQR: 41.5), as were final decision times (community review: median 138, IQR: 22.5; standard review: median 211, IQR: 166.0). A survey of the Live Review attendees (n = 13; 30% response rate) on a scale of 1 'Highly Disagree' to 5 'Highly Agree' showed median 'Agree' to 'Highly Agree' scores on several questions including the review being respectful, time efficient and scientifically rigorous (median scores: 5, 4, 4, respectively). The innovative Live Review approach was as efficient as the standard review process in the journal and was rated positively by those surveyed. The small sample size inherent to exploratory studies limits conclusions generalizing to larger sample sizes. We discuss how live, community-based review approaches could be further developed, scaled and sustained.

Convolutional neural networks (CNNs) provide powerful models of neural sensory encoding, but their complexity makes it difficult to discern computations that support their performance. Here, to address this limitation, we developed a linear–nonlinear subspace model that identifies the most informative sensory dimensions captured by a CNN. A CNN was trained on single-neuron data recorded from auditory cortex of ferrets during presentation of a large natural sound set. Each neuron’s linear tuning subspace was computed by applying dimensionality reduction to the gradient of CNN output relative to input. Subspace projections were combined nonlinearly to predict neural activity. The resulting model was functionally equivalent to the CNN. Analysis of trained models showed that responses of local neural populations sparsely tiled a shared stimulus subspace. Encoding properties also differed between cell types and layers, reflecting their position in the cortical circuit. More generally, these results establish a framework for interpreting deep-learning-based encoding models.

Academic peer review is fundamental for scientific knowledge dissemination, and various initiatives are exploring how the peer-review process could be more open, efficient and rewarding. We report five case studies where a live community-based review session was integrated into the editorial workflow of an academic journal (Current Research in Neurobiology; CRNEUR). Five manuscripts, submitted as preprints, underwent Live Review—a structured collaborative review session led by PREreview, an open science project advancing openness in scholarly evaluation. With each case, PREreview team members facilitated a 90-minute online discussion where registered participants provided real-time discussion and worked together on an online structured peer-review document. Authors could join as observers or to answer questions, and journal editors could join as observers. Participants then volunteered to write up the session notes into a final review and summary statement. Review participants had the option to sign the review. The finalized review was then published on PREreview’s open preprint review platform approximately two weeks after the Live Review session. The published review was assigned a Digital Object Identifier (DOI) for participating reviewers to obtain credit for their reviewing effort. The published review was then incorporated into CRNEUR’s editorial process to inform editorial decisions. Results suggest that the speed of this community review can be as rapid as the standard peer-review process for CRNEUR during the same time period, and a small sample size survey of the Live Review pilots attendees showed agreement on several questions including the review being respectful, time efficient and scientifically rigorous. We discuss how live, community-based review approaches could be further developed, scaled and sustained.

Much remains unknown about the computations that allow animals to flexibly integrate across multiple timescales in natural sounds. One key question is whether multiscale integration is accomplished by diverse populations of neurons, each of which integrates information within a constrained temporal window, or whether individual units effectively integrate across many different temporal scales depending on the information rate. Here, we show that responses from neurons throughout the ferret auditory cortex are nearly completely unaffected by sounds falling beyond a time-limited “integration window”. This window varies substantially across cells within the auditory cortex (∼15 to ∼150 ms), increasing substantially from primary to non-primary auditory cortex across all cortical layers, but is unaffected by the information rate of sound. These results indicate that multiscale computation is predominantly accomplished by diverse and hierarchically organized neural populations, each of which integrates information within a highly constrained temporal window.

In everyday hearing, listeners face the challenge of understanding behaviorally relevant foreground stimuli (speech, vocalizations) in complex backgrounds (environmental, mechanical noise). Prior studies have shown that high-order areas of human auditory cortex (AC) preattentively form an enhanced representation of foreground stimuli in the presence of background noise. This enhancement requires identifying and grouping the features that comprise the background so they can be removed from the foreground representation. To study the cortical computations supporting this process, we recorded single-unit activity in AC of male and female ferrets during the presentation of concurrent natural sounds from these two categories. In contrast to expectations from studies in high-order AC, single-unit responses to foreground sounds were strongly reduced relative to the paired background in primary and secondary AC. The degree of reduction could not be explained by a neuron's preference for the foreground or background stimulus in isolation but could be partially explained by spectrotemporal statistics that distinguish foreground and background categories. Responses to synthesized sounds with statistics either matched or randomized relative to natural sounds showed progressively decreased reduction of foreground responses as natural sound statistics were removed. These results challenge the expectation that cortical foreground representations emerge directly from a mixed representation in the auditory periphery. Instead, they suggest the early AC maintains a robust representation of background noise. Strong background representations may produce a distributed code, facilitating selection of foreground signals from a relatively small subpopulation of AC neurons at later processing stages.

Convolutional neural networks (CNNs) provide powerful models of neural sensory encoding, but their complexity makes it difficult to discern computations that support their performance. To address this limitation, we developed a linear-nonlinear subspace model that identifies the most informative sensory dimensions captured by a CNN. A CNN was trained on single-neuron data recorded from auditory cortex of ferrets during presentation of a large natural sound set. Each neuron’s linear tuning subspace was computed by applying dimensionality reduction to the gradient of CNN output relative to input. Subspace projections were combined nonlinearly to predict neural activity. The resulting model was functionally equivalent to the CNN. Analysis of trained models showed that responses of local neural populations sparsely tiled a shared stimulus subspace. Encoding properties also differed between cell types and layers, reflecting their position in the cortical circuit. More generally, these results establish a framework for interpreting deep learning-based encoding models. Significance statement: Auditory cortex mediates the representation and discrimination of complex sound features. Many models have been proposed for cortical sound encoding, varying in their generality, interpretability, and ease of fitting. It has been difficult to determine if and what different functional properties are captured by different models. This study shows that two families of encoding models, convolutional neural networks (CNNs) and tuning subspace models, account for the same functional properties, providing an important analytical link between accurate models that are easy to fit (CNNs) and models that are straightforward to interpret (tuning subspace).

Categorical sensory representations are critical for many behaviors, including speech perception. In the auditory system, categorical information is thought to arise hierarchically, becoming increasingly prominent in higher-order cortical regions. The neural mechanisms that support this robust and flexible computation remain poorly understood. Here, we studied sound representations in the ferret primary and non-primary auditory cortex while animals engaged in a challenging sound discrimination task. Population-level decoding of simultaneously recorded single neurons revealed that task engagement caused categorical sound representations to emerge in non-primary auditory cortex. In primary auditory cortex, task engagement caused a general enhancement of sound decoding that was not specific to task-relevant categories. These findings are consistent with mixed selectivity models of neural disentanglement, in which early sensory regions build an overcomplete representation of the world and allow neurons in downstream brain regions to flexibly and selectively read out behaviorally relevant, categorical information.

Categorical sensory representations are critical for many behaviors, including speech perception. In the auditory system, categorical information is thought to arise hierarchically, becoming increasingly prominent in higher order cortical regions. The neural mechanisms that support this robust and flexible computation remain poorly understood. Here, we studied sound representations in primary and non-primary auditory cortex while animals engaged in a challenging sound discrimination task. Population-level decoding of simultaneously recorded single neurons revealed that task engagement caused categorical sound representations to emerge in non-primary auditory cortex. In primary auditory cortex, task engagement caused a general enhancement of sound decoding that was not specific to task-relevant categories. These findings are consistent with mixed selectivity models of neural disentanglement, in which early sensory regions build an overcomplete representation of the world and allow neurons in downstream brain regions to flexibly and selectively read out behaviorally relevant, categorical information.

In everyday hearing, listeners encounter complex auditory scenes containing overlapping sounds that must be grouped into meaningful sources, or streamed, to be perceived accurately. A common example of this problem is the perception of a behaviorally relevant foreground stimulus (speech, vocalizations) in complex background noise (environmental, machine noise). Studies using a foreground/background contrast have shown that high-order areas of auditory cortex in humans pre-attentively form an enhanced representation of the foreground over background stimulus. Achieving this invariant foreground representation requires identifying and grouping the features that comprise the background noise so that they can be removed from the representation of the foreground. To study the cortical computations underlying representation of concurrent background (BG) and foreground (FG) stimuli, we recorded single unit responses in the auditory cortex (AC) of ferrets during presentation of natural sound excerpts from these two categories. In primary and secondary AC, we found overall suppression of responses when BGs and FGs were presented concurrently relative to the sum of responses to the same stimuli in isolation. Surprisingly, and in contrast to percepts that emphasize dynamic FGs, responses to FG sounds were suppressed relative to the paired BG sound. The degree of relative FG suppression could be explained by spectro-temporal statistics unique to each natural sound. Moreover, systematic degradation of the same spectro-temporal features decreased FG suppression as the sound categories became progressively less statistically distinct. The strongly suppressed representation of FG sounds in single units of AC in the presence of BG sound reveals a novel insight into how complex acoustic scenes are encoded at early stages of auditory processing.