LANE McINTOSH

Deep Learning Models of the Retina

A central challenge in sensory neuroscience is to understand neural computations and circuit mechanisms that underlie the encoding of ethologically relevant, natural stimuli. In neural circuits, ubiquitous nonlinear processes present a significant obstacle to the creation of accurate computational models of responses to natural stimuli. We demonstrate that deep convolutional neural networks capture retinal responses to natural scenes nearly to within the variability of a cell's response, and are markedly more accurate than previous models. We are then able to probe the learned models to gain insights about the retina, for instance how it compresses natural scenes efficiently through feedforward inhibition and how it transforms potentially large sources of extrinsic and intrinsic noise into sub-Poisson variability. Overall, this work demonstrates that CNNs not only accurately capture sensory circuit responses to natural scenes, but also can yield information about the circuit's internal structure and function.
Lane McIntosh*, Niru Maheswaranathan*, Aran Nayebi, Surya Ganguli, Stephen Baccus
Accepted Paper, Advances in Neural Information Processing Systems (NIPS), 2016
Accepted Talk, Society for Neuroscience, 2016
Accepted Poster, Computational and Systems Neuroscience (COSYNE), 2016
NVIDIA Best Poster, SCIEN Industry Affiliates Meeting (image processing), 2015
Top 10% Poster Award, CS231n Convolutional Neural Networks, 2015

NIPS 2016 paper COSYNE 2016 Poster

Synchronous inhibitory pathways create both efficiency and diversity in the retina

Retinal ganglion cells, the bottleneck of all visual information to the brain, have linear response properties that appear to maximize information between the visual world and the retinal ganglion cell responses, subject to a variance constraint. In this paper I contribute a new theoretical finding that generating the ganglion cells' linear receptive field from inhibitory interneurons with disparate spatial scales provides a basis that allows the receptive field to maximize information under a wide range of environments whose signal-to-noise ratios vary by orders of magnitude.
Mihai Manu*, Lane McIntosh*, David Kastner, Benjamin Naecker, and Stephen Baccus
In Review, Nature Neuroscience 2017

Video-based Event Recognition

How can we automatically extract events from video? We used a database of surveillance videos and examined the performance of SVMs and Convolutional Neural Networks in detecting events like people getting in and out of cars.
Ian Ballard* and Lane McIntosh*
CS221 Artificial Intelligence Poster, 2014

Learning Predictive Filters

How should an intelligent system intent on only keeping information predictive about the future filter its data? We analytically find the optimal predictive filter for Gaussian input using recent theorems from the information bottleneck literature. Using numerical methods, we then show the resemblance of these optimally predictive filters to the receptive fields in early visual pathways of vertebrates.
Lane McIntosh
CS229 Machine Learning Poster, 2013

Thermodynamics of Prediction in Model Neurons

Recent theorems in nonequlibrium thermodynamics show that information processing inefficiency provides a lower bound for energy dissipation in certain systems. We extend these results to model neurons and find that adapting neurons that match the timescale of their inputs perform predictive inference while minimizing energy inefficiency.
Lane McIntosh and Susanne Still
Master's Thesis, 2012