Perceptual Systems Based on Cortical Computation
Abstract
Statement of Work: This project will encompass the following work: 1) Incorporate properly inferred sparse representations at each stage of deep neural network processing pipeline; 2) Utilize neurally plausible pose normalized representations; Replace sum pooling or maxpooling operation with multiplicatively gated connections, and infer control neuron activities by maximizing match between input image and memory. 3) Formulate methods for resolving ambiguity at lower levels of representation by using top-down feedback; Use Hierarchical Bayesian Inference framework to infer representations at each stage through a combination of bottom-up (data) and top-down (model) signals. 4) Create brain-inspired video processing models; Replace spatial filters with spatiotemporal filters, and use factorized representations to separate form and motion. This is a fundamental research project that is not expected to produce any developmental items. Should any developmental items result from this work they will have both civilian and military applications. Objective: This project will advance the state of the art in building robust, perceptual systems by leveraging recent advances in machine learning together with insights from neuroscience regarding cortical computation. Over the past several years, the field of machine learning has undergone a dramatic advance due to the advent of deep net architectures, which use multiple stages of neural-like computational elements to process sensory input. These systems now routinely achieve the best performance at object recognition, speech recognition, and other benchmark classification tasks. However, despite the fact that these architectures are inspired by the hierarchical architecture of mammalian cortex, they are lacking many of the properties of biological systems which are thought to be important to providing robust performance: sparse representation, processing of temporal information, and mechanisms for attentional selection and top-down feedback. Here, we seek to incorporate these aspects of cortical computation into deep net architectures, with the aim of achieving an improvement in efficiency and performance. This a one year pilot study that seeks a proof of concept demonstration. Approach: The approach is primarily based on algorithm development and testing. Our team possesses expertise in sparse coding models, unsupervised learning algorithms, and associative memory models, especially biophysically realistic mechanistic models of these processes. We also have experience in developing deployable vision systems based on efficient large scale algorithms and infrastructures, specifically the widely adopted open source deep learning software framework developed at Berkeley, CAFFE 1, which will form the software foundation for this project. We will conduct empirical evaluations of our proposed brain-inspired innovations, measuring the impact on performance using large-scale image detection and fine-grained recognition tasks on established benchmarks. Overall Merit and ONR Mission/Relevance: This project will advance the state of the art in autonomous systems. It will pave the way for building machines that can perceive and act in complex environments, which is central to saving lives in the battlefield, rescue operations, or other hazardous working conditions.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 12, 2016
- Source ID
- N000141512731
Entities
People
- Bruno A. Olshausen
Organizations
- Office of Naval Research
- United States Navy
- University of California Regents