Deep neural network models of sensory systems: windows onto the role of task constraints

Kell, Alexander J.E.; McDermott, Josh H.

doi:10.1016/j.conb.2019.02.003

Cited by 89 publications

(87 citation statements)

References 66 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We first test whether a pre-trained VGG-16 model exhibits the classical oblique effect, assessed using the Fisher information measured at entire layers of the network, and the distribution of single-unit tuning properties. In addition to a test of the efficient coding hypothesis, measuring orientation bias in this pre-trained model will provide an assessment of whether existing CNNs, often used as models of the primate visual system (18,19), exhibit this defining characteristic of biological vision.…”

Section: Introductionmentioning

confidence: 99%

Biased orientation representations can be explained by experience with non-uniform training set statistics

Henderson¹,

Serences²

2020

Preprint

View full text Add to dashboard Cite

Visual acuity is better for vertical and horizontal compared to other orientations. This cross-species phenomenon is often explained by "efficient coding", whereby more neurons show sharper tuning for the orientations most common in natural vision. However, it is unclear if experience alone can account for such biases. Here, we measured orientation representations in a convolutional neural network, VGG-16, trained on modified versions of ImageNet (rotated by 0, 22.5, or 45 degrees counter-clockwise of upright). Discriminability for each model was highest near the orientations that were most common in the network's training set. Furthermore, there was an over-representation of narrowly tuned units selective for the most common orientations. These effects emerged in middle layers and increased with depth in the network. Our results suggest that biased orientation representations can emerge through experience with a non-uniform distribution of orientations. These findings thus support the efficient coding hypothesis and highlight that biased training data can systematically distort processing in CNNs.

show abstract

Section: Introductionmentioning

confidence: 99%

Biased orientation representations can be explained by experience with non-uniform training set statistics

Henderson¹,

Serences²

2020

Preprint

View full text Add to dashboard Cite

show abstract

“…The UoI framework is naturally extendable to such methods, including column subset selection [53] and non-negative matrix factorization [54]. Furthermore, recent work has found success in using artificial neural networks (ANNs) to model neural activity, which excel at predictive performance [113]. Since ANNs are highly parameterized, these models are not interpretable in the sense that their parameter values do not directly convey neuroscientific meaning.…”

Section: Discussionmentioning

confidence: 99%

Accurate inference in parametric models reshapes neuroscientific interpretation and improves data-driven discovery

Sachdeva

Livezey

Dougherty

et al. 2020

Preprint

View full text Add to dashboard Cite

A central goal of systems neuroscience is to understand the relationships amongst constituent units in neural populations and their modulation by external factors using high-dimensional and stochastic neural recordings. Statistical models, particularly parametric models, play an instrumental role in accomplishing this goal, because their fitted parameters can provide insight into the underlying biological processes that generated the data. However, extracting conclusions from a parametric model requires that it is fit using an inference procedure capable of selecting the correct parameters and properly estimating their values. Traditional approaches to parameter inference have been shown to suffer from failures in both selection and estimation. Recent development of algorithms that ameliorate these deficiencies raises the question of whether past work relying on such inference procedures have produced inaccurate systems neuroscience models, thereby impairing their interpretation. Here, we used the Union of Intersections, a statistical inference framework capable of state-of-the-art selection and estimation performance, to fit functional coupling, encoding, and decoding models across a battery of neural datasets. We found that, compared to baseline procedures, UoI inferred models with increased sparsity, improved stability, and qualitatively different parameter distributions, while maintaining predictive performance across recording modality, brain region, and task. Specifically, we obtained highly sparse functional coupling networks with substantially different community structure, more parsimonious encoding models, and decoding models that rely on fewer single-units. Together, these results demonstrate that accurate parameter inference reshapes interpretation in diverse neuroscience contexts. The ubiquity of model-based data-driven discovery in biology suggests that analogous results would be seen in other fields.

show abstract

“…This model also sets a precedent for how traditional approaches from computational neuroscience can be incorporated with the increasingly popular approach of using deep neural networks to study the brain (Kell and McDermott, 2019;Lindsay, 2020;Yamins and DiCarlo, 2016).…”

Section: Discussionmentioning

confidence: 99%

A unified circuit model of attention: Neural and behavioral effects

Lindsay

Rubin

Miller

2019

Preprint

View full text Add to dashboard Cite

Selective visual attention modulates neural activity in the visual system and leads to enhanced performance on difficult visual tasks. Here, we use an existing circuit model of visual cortex, known as the stabilized supralinear network, to demonstrate that many neural correlates of attention can arise from simple circuit mechanisms. Using different variants of the model we replicate results from studies of both feature and spatial attention. In addition to firing rate changes, we also replicate findings regarding how attention impacts trial-to-trial variability. Finally, we expand this circuit model into an architecture that can perform visual tasks in order to show that these neural effects can enhance detection performance. This work advances our understanding of the physical underpinnings of attention.

show abstract

Deep neural network models of sensory systems: windows onto the role of task constraints

Cited by 89 publications

References 66 publications

Biased orientation representations can be explained by experience with non-uniform training set statistics

Biased orientation representations can be explained by experience with non-uniform training set statistics

Accurate inference in parametric models reshapes neuroscientific interpretation and improves data-driven discovery

A unified circuit model of attention: Neural and behavioral effects

Contact Info

Product

Resources

About