Bruno A. Olshausen - US grants

DESCRIPTION (Adapted from applicant's abstract): A major limitation to our understanding of visual cortical function is the lack of computational theories capable of making useful, testable predictions for what the cortex should be doing. The purpose of this study is to investigate what may be learned about information processing in visual cortex from efficient coding principles. Methods will be developed for representing the structure in images based on probabilistic inference, and these will be related to known neurobiological substrates in a detailed manner in order to make predictions about visual cortical function. Understanding how the cortex processes visual information is an important step in developing therapies for patients who have lost aspects of visual function due to cortical damage, as well as in the development of visual prostheses capable of providing appropriate cortical stimulation from artificial vision devices. The aspects of visual cortical function that the study aims to shed light on are the properties of feature selectivity, form-invariance, and the role of feedback connections in shaping neural response properties and in mediating visual perception. These issues will be addressed as part of five specific aims. The first is to develop a functional model of the horizontal connections in area V1 based on the statistical structure of natural images. This model will be related to the structure of long-range horizontal fibers in order to make predictions about the role of this form of feedback within V1. The second aim is to develop a model neural system capable of learning the structure of objects independent of variations in position, size, or other geometric transformations. The model will be used to help understand how form-invariance is established in cortical neurons. The third aim is to formulate a model of occlusion in images, which will be used to shed light on how figure-ground segregation could be performed by cortical mechanisms. The fourth aim is to develop a functional model of top-down cortical feedback based on a hierarchical image model. The existence of such a system that utilizes top-down feedback to solve practical problems in vision will help to elucidate a possible role for two-way information processing in the cortical hierarchy. The fifth aim is to test these models through psychophysical experiments. The results of these studies will lead to advances in our understanding of information processing in visual cortex, and possibly shed light on the nature of cortical information processing in general.

This project investigates the Next Generation Network Technology and Systems capable of understanding and learning the high-level perspective of the network. The proposed approach pursues a new cognitive intelligent networking paradigm that maintains the success of today's Internet but which also incorporates cognitive intelligence in the network--a new networking technique that provides the ability for the network to know what it is being asked to do, so that it can step-by-step take care of itself as it learns more. In particular, we explore new networking architecture and network elements that will lead to a future network with (a) improved robustness and adaptability, (b) improved usability and comprehensibility, (c) improved security and stability, and (d) reduced human intervention for operation and configuration. This project pursues a set of comprehensive studies that seek innovations through the design and modeling of a new brain-reflex cognitive intelligence architecture, an intelligent programmable network elements architecture, and an intelligent network control and management design.

Broader Impact: The team approach covering neuroscience, datamining, computer science, systems engineering, artificial intelligence, and networking will provide rich opportunities for students to learn beyond their primary fields of study. New courses developed by the faculty members will disseminate the new material covering neuroscience and information technology.

Proposal No: 0749049
PI: Friedrich T. Sommer

Award Abstract:

This award supports services and infrastructure for sharing of computational neuroscience data as part of an exploratory activity aimed at catalyzing rapid and innovative advances in computational neuroscience and related fields. The core facility will provide transparent access to shared resources in a manner that scales up to large data sets. Services will be designed to lessen the burden on contributors to make their data or other resources available and to optimize the ability of the user community to identify and use those resources. Community- and market-oriented mechanisms will be developed to identify resources of particular significance for the field, and to solicit feedback from relevant communities. It is anticipated that the availability of high quality data will offer unprecedented opportunities for new types of discoveries, development of new methods, and development of new interdisciplinary collaborations. This new activity will also assist and drive teaching in computational neuroscience, through the exchange of datasets, stimuli, and analysis and modeling tools among modelers, experimentalists, and students.

Project Summary/Abstract Using as a starting point the postulate that sensory systems have evolved to perform optimal transformations on behaviorally relevant or natural stimuli, we are using signal analysis methods and information theoretic principles to develop a theory of auditory processing. The purpose of our theory is not just to describe but to understand the neural representation of acoustic communication signals, including speech and music. First, we plan on analyzing the statistics of natural sounds and of speech, music and birdsong in particular. We propose to search for theoretical representations of sounds based on principles of statistical independence and sparse representation. Our derived representations will also attempt to maximize differences between acoustic features that meditate the qualitatively different acoustical percepts of rhythm, timbre and pitch. Second, we will test the validity of these theoretically derived representations in psychophysical experiments in humans, and behavioral experiments in songbirds. These experiments will test the effect on perception of systematically removing acoustic features along the particular dimensions that were derived in the statistical analysis. Third, we will develop information theoretic tools that will allow us to estimate the amount of redundancy in a neuronal ensemble response. These measures will be used to quantify how the neural representation changes as one ascends the auditory processing stream and to test whether the neural representation is becoming more sparse and independent as we theorized. Finally, we will record the neural responses in primary and secondary auditory areas in songbirds to playback of song and filtered song. The data from these neurophysiological experiments will be used to: 1) test the utility of the statistically derived representations to predict responses of single auditory neurons, 2) correlate neural responses and behavioral responses, 3) assess the nature of non-linearities in the response, and 4) test the assumptions of independence at the ensemble level. Our studies will give us insight on how speech, music and other complex sounds are processed by the auditory system. These studies could be instrumental in the development of novel methods for sound processing for hearing aids and auditory neural prosthetics, as well as diagnostic tools for classifying language and learning disorders.

This project aims to advance neural data analysis and image processing by exploiting the structure in multivariate phase representations. Combining insights from neural computation with advances in multivariate statistics, mathematical signal analysis, and machine learning, the project aims to build multivariate statistical models of angular variables that capture the dependencies between complex and hypercomplex phase variables. Recursive estimation techniques will be developed to allow for optimal estimation of distributions from noisy data and prediction of their temporal evolution. The models developed in this proposal will be applied to current problems in neuroscience and image processing.

As a foundation for the application domains, a recently developed method for estimating the parameters of stationary multivariate phase distributions will be generalized to situations in which the parameters are time-varying and the measurements are noisy, linear mixtures of the underlying sources. A recursive filtering model, similar to the classical Kalman filter, will be developed, that produces an optimal online estimate of latent phase variables in response to a sequence of noisy measurements.

In the first application domain, this model will be used to infer connectivity and temporal interactions among populations of neural oscillators from physiological measurements. The model will also be used to detect transient changes in connectivity by utilizing a mixture model of phase dynamics. These estimation techniques will be evaluated on simulated data and then applied to the analysis of neurophysiological recordings to better elucidate network dynamics.

In the application domain of natural image statistics, the model will be extended to handle hypercomplex phase variables in order to model the phase and orientation of edges in natural images, which contain rich information that may be exploited for image analysis and object recognition. The project aims to develop a model that describes the spatial dependencies among these variables as a nonlinear mixture of multivariate phase distributions.

https://redwood.berkeley.edu/wiki/NSF_Funded_Research

How does a vision system recover the 3-dimensional structure of the world -- such as the layout of the environment, surface shape, or object motion -- from the dynamic 2-dimensional images received by the sensors in a camera, or the retinas in our eyes? This problem is fundamental to both computer and biological vision. Computer vision has developed a variety of algorithms for estimating specific aspects of a scene such as the 3-dimensional positions of points whose correspondence over time can be established, but obtaining complete and robust scene representations for complex natural scenes and viewing conditions remains a challenge. Biological vision systems have evolved impressive capabilities that suggest they have detailed and robust representations of the 3-dimensional world, but the neural representations that subserve this are poorly understood and neurophysiological studies thus far have provided little insight into the computational process. This project will pursue an interdisciplinary approach by attempting the understand the universal principles that lie at the heart of 3-dimensional scene analysis.

Specifically, the project will 1) develop a novel class of computational models that recover and represent 3-dimensional scene information, 2) collect high quality video and range data of dynamic natural scenes under a variety of controlled motion conditions, and 3) test the perceptual implications of these models in psychophysical experiments. The computational models will utilize non-linear decomposition - i.e., the ability to explain complex, time-varying images in terms of the non-linear interaction of multiple factors, such as the interaction between observer motion, the 3-dimensional scene layout, and surface patterns. Importantly, the components of these models will be adapted to the statistics of natural motion patterns that arise from observer motion through natural scenes and movement around points of fixation.

The project is a collaboration between three laboratories that have played a leading role in developing theoretical models of natural image statistics, visual neural representations, and perceptual processes. The investigators seek to combine their efforts to develop new models, data sets, and characterizations of 3-dimensional natural scene structure that go beyond previous studies of natural image statistics, and that can be tested in neurophysiological and psychophysical experiments. This project has the potential to bring about fundamental advances in neuroscience, visual perception, and computer vision by developing new classes of models that robustly infer representations of the 3-dimensional natural environment. It will create a set of high quality databases that will be made available to help other investigators study these issues. It will also open up new possibilities for generating realistic stimuli that can guide novel investigations of neural representation and processing.

The project will aim to develop computing hardware and software that improve the energy efficiency of learning machines by many orders of magnitude. In doing so it will enable large societal adoption of such machines, paving the way for new applications in diverse areas such as manufacturing, healthcare, agriculture, and many others. For example, machines that learn the behavioral trends of individual human beings by collecting data from myriads of sensors may be able to design the most appropriate drugs. Similarly, one may envision machines that learn trends in the weather and thereby assist in predicting the most optimized preparations for the next crop cycle. The possibilities are literally endless. However, the canonical learning machines of today need huge amount of energy, significantly hindering their adoption for widespread applications. The goal of this project will be to explore, evaluate and innovate new hardware and software paradigms that could reduce energy dissipation in learning machines by a significant amount. The team of researchers consists of experts in mathematics, neuroscience, electronic devices and materials and computer circuit and system design that will foster a unique platform for both innovative research and interdisciplinary training of graduate students.

We are witnessing a regimental shift in the computing paradigm. For a vast number of applications, cognitive functions such as classification, recognition, synthesis, decision-making and learning are gaining rapid importance in a world that is infused with sensing modalities, often paraphrased under a common term of "Big Data", that are in critical need of efficient information-extraction. This is in sharp contrast to the past when the central objective of computing was to perform calculations on numbers and produce results with extreme numerical accuracy. We aim to approach this problem by exploiting cognitive models that have shown efficacy in "one shot" learning. In this approach, the information is represented by means of high dimensional (HD) vectors. These vectors follow a set of predetermined mathematical operations that ensure that the resulting vector after such operations is unique. The uniqueness can in turn be used as "learning" and the predefined nature of mathematical operations make the learning "one shot". When paired with traditional artificial neural network or deep learning algorithms, such "one shot" learning could significantly reduce the number of necessary computing operations, leading to orders of magnitude reduction in energy dissipation. We shall explore the entire computer hierarchy, staring from materials and devices, all the way up to system design and optimization to exploit the unique capabilities afforded by the HD computing, with the ultimate objective of realizing energy efficient learning machines (ENIGMA).

In the modern era of big data, a crucial challenge is to discover useful information that is buried in highly redundant, seemingly irrelevant, incomplete, or even corrupted data sets. Such information is often contained in certain low-dimensional structures hidden within the high-dimensional space of the data, or may only depend on a small subset of the data. How to extract this information efficiently and automatically remains an open problem. This project brings together two emerging areas of research — hyperdimensional (HD) computing and geometric algebra (GA) — to tackle this problem from a new stand point by investigating the data representation and the intrinsic geometry of the data. This research is also the first in a systematic quest to uncover the potential of using the high-dimensional generalization of complex numbers in analyzing and discovering patterns in large-scale sensing data. The success of this research can help advance the capability of other machine learning models, such as deep neural networks, which are mostly based on real numbers today. It also brings a powerful mathematical tool (GA) which is mainly known in the physics community into the machine learning community.

HD computing is a brain-inspired framework for machine learning and artificial intelligence that is based on representing quantities or symbols as high-dimensional vectors and manipulating vectors with simple operations. In recent work by the investigators, it was shown that by using complex-valued vectors in HD computing it is possible to encode images in such a way that patterns can be effectively recognized by a factorization of HD vectors. To build on this direction, they are exploring the use of geometric algebras which generalize complex numbers to any n-dimensional space. The following thrusts form the core of this research: (1) explore ways of mapping data into the geometric algebra space; (2) investigate how to integrate geometric algebra with the operations of HD computing; (3) apply these methods to real application domains such as multi-microphone speech recognition or distributed sensing to evaluate their efficacy and computational efficiency.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.