skip to main content

This content will become publicly available on January 1, 2023

Title: Emergence of direction-selective retinal cell types in task-optimized deep learning models
Convolutional neural networks (CNNs), a class of deep learning models, have experienced recent success in modeling sensory cortices and retinal circuits through optimizing performance on machine learning tasks, otherwise known as task optimization. Previous research has shown task-optimized CNNs to be capable of providing explanations as to why the retina efficiently encodes natural stimuli and how certain retinal cell types are involved in efficient encoding. In our work, we sought to use task-optimized CNNs as a means of explaining computational mechanisms responsible for motion-selective retinal circuits. We designed a biologically constrained CNN and optimized its performance on a motion-classification task. We drew inspiration from psychophysics, deep learning, and systems neuroscience literature to develop a toolbox of methods to reverse engineer the computational mechanisms learned in our model. Through reverse engineering our model, we proposed a computational mechanism in which direction-selective ganglion cells and starburst amacrine cells, both experimentally observed retinal cell types, emerge in our model to discriminate among moving stimuli. This emergence suggests that direction-selective circuits in the retina are ecologically designed to robustly discriminate among moving stimuli. Our results and methods also provide a framework for how to build more interpretable deep learning models and how to understand them.
Authors:
; ;
Award ID(s):
1810758 2003830
Publication Date:
NSF-PAR ID:
10324411
Journal Name:
Journal of computational biology
Volume:
29
Issue:
4
Page Range or eLocation-ID:
370-381
ISSN:
1066-5277
Sponsoring Org:
National Science Foundation
More Like this
  1. Convolutional neural networks (CNN) are an emerging technique in modeling neural circuits and have been shown to converge to biologically plausible functionality in cortical circuits via task-optimization. This functionality has not been observed in CNN models of retinal circuits via task-optimization. We sought to observe this convergence in retinal circuits by designing a biologically inspired CNN model of a motion-detection retinal circuit and optimizing it to solve a motion-classification task. The learned weights and parameters indicated that the CNN converged to direction-sensitive ganglion and amacrine cells, cell types that have been observed in biology, and provided evidence that task-optimization ismore »a fair method of building retinal models. The analysis used to understand the functionality of our CNN also indicates that biologically constrained deep learning models are easier to reason about their underlying mechanisms than traditional deep learning models.« less
  2. Convolutional neural networks (CNN) are an emerging technique in modeling neural circuits and have been shown to converge to biologically plausible functionality in cortical circuits via task-optimization. This functionality has not been observed in CNN models of retinal circuits via task-optimization. We sought to observe this convergence in retinal circuits by designing a biologically inspired CNN model of a motion-detection retinal circuit and optimizing it to solve a motion-classification task. The learned weights and parameters indicated that the CNN converged to direction-sensitive ganglion and amacrine cells, cell types that have been observed in biology, and provided evidence that task-optimization ismore »a fair method of building retinal models. The analysis used to understand the functionality of our CNN also indicates that biologically constrained deep learning models are easier to reason about their underlying mechanisms than traditional deep learning models.« less
  3. The optic nerve transmits visual information to the brain as trains of discrete events, a low-power, low-bandwidth communication channel also exploited by silicon retina cameras. Extracting highfidelity visual input from retinal event trains is thus a key challenge for both computational neuroscience and neuromorphic engineering. Here, we investigate whether sparse coding can enable the reconstruction of high-fidelity images and video from retinal event trains. Our approach is analogous to compressive sensing, in which only a random subset of pixels are transmitted and the missing information is estimated via inference. We employed a variant of the Locally Competitive Algorithm to infermore »sparse representations from retinal event trains, using a dictionary of convolutional features optimized via stochastic gradient descent and trained in an unsupervised manner using a local Hebbian learning rule with momentum. We used an anatomically realistic retinal model with stochastic graded release from cones and bipolar cells to encode thumbnail images as spike trains arising from ON and OFF retinal ganglion cells. The spikes from each model ganglion cell were summed over a 32 msec time window, yielding a noisy rate-coded image. Analogous to how the primary visual cortex is postulated to infer features from noisy spike trains arising from the optic nerve, we inferred a higher-fidelity sparse reconstruction from the noisy rate-coded image using a convolutional dictionary trained on the original CIFAR10 database. To investigate whether a similar approachworks on non-stochastic data, we demonstrate that the same procedure can be used to reconstruct high-frequency video from the asynchronous events arising from a silicon retina camera moving through a laboratory environment.« less
  4. Regenerative retinal therapies have introduced progenitor cells to replace dysfunctional or injured neurons and regain visual function. While contemporary cell replacement therapies have delivered retinal progenitor cells (RPCs) within customized biomaterials to promote viability and enable transplantation, outcomes have been severely limited by the misdirected and/or insufficient migration of transplanted cells. RPCs must achieve appropriate spatial and functional positioning in host retina, collectively, to restore vision, whereas movement of clustered cells differs substantially from the single cell migration studied in classical chemotaxis models. Defining how RPCs interact with each other, neighboring cell types and surrounding extracellular matrixes are critical tomore »our understanding of retinogenesis and the development of effective, cell-based approaches to retinal replacement. The current article describes a new bio-engineering approach to investigate the migratory responses of innate collections of RPCs upon extracellular substrates by combining microfluidics with the well-established invertebrate model of Drosophila melanogaster. Experiments utilized microfluidics to investigate how the composition, size, and adhesion of RPC clusters on defined extracellular substrates affected migration to exogenous chemotactic signaling. Results demonstrated that retinal cluster size and composition influenced RPC clustering upon extracellular substrates of concanavalin (Con-A), Laminin (LM), and poly-L-lysine (PLL), and that RPC cluster size greatly altered collective migratory responses to signaling from Fibroblast Growth Factor (FGF), a primary chemotactic agent in Drosophila. These results highlight the significance of examining collective cell-biomaterial interactions on bio-substrates of emerging biomaterials to aid directional migration of transplanted cells. Our approach further introduces the benefits of pairing genetically controlled models with experimentally controlled microenvironments to advance cell replacement therapies.« less
  5. Two interleaved stimulus sets were identical except for the background. In one, the flow stimuli background was the mid-gray of the interstimulus interval (equal background, eqbg), leading to a change of 9-10% in the space-average luminance. In the other, the space-average luminance of the entire stimulus field was adjusted to a constant (equal luminance, eqlum) within 0.5%; i.e., the background was slightly lightened when the dots in the flow were dark, and darkened when the dots were bright. Most cortical cells appeared to respond similarly to the two stimulus sets, as if stimulus structure mattered but not the background change,more »while the responses of most retinal ganglion cells appeared to differ between the two conditions. Machine learning algorithms confirmed this quantitatively. A manifold embedding of neurons to the two stimulus sets was constructed using diffusion maps. In this manifold, the responses of the same cell to eqlum and eqbg stimuli were significantly closer to one another for V1 rather than for the retina. Geometrically, the median ratio of the distance between the responses of each cell to the two stimulus sets as compared to the distance to the closest cell on the manifold was 3.5 for V1 compared to 12.7 for retina. Topologically, the fraction of cells for which the responses of the same cell to the two stimulus sets were connected in the diffusion map datagraph was 53% for V1 but only 9% for retina; when retina and cortex were co-embedded in the manifold, these fractions were 44% and 6%. While retina and cortex differ on average, it will be intriguing to determine whether particular classes of retinal cells behave more like V1 neurons, and vice versa.« less