Lectures 
Building and Interpreting Populations of Model Visual Cortical Neurons 
Bert Shi, Hong Kong University of Science and Technology 

We describe multi-dimensional selectivity in primary visual cortical neurons: the concept that the response of these neurons is tuned to respond maximally along a variety of visual dimensions including position, spatial/temporal frequency, orientation, disparity and motion. We then describe the approach and motivation for the choices of what and how to model this multi-dimensional selectivity. We then describe and compare two neuromorphic systems we have implemented. The first consists of multiple chips which communicate via spikes transmitted using the AER protocol. Internally, the chips process signals using locally connected networks. We conclude that we now have the technology to build large networks of such neurons. At the same time, because less work has done in modelling the processing in areas beyond the primary visual cortex, these models are in a state of flux. Thus, we need to develop a more rapidly reconfigurable hardware platform that support real-time implementation of such neurons, but that is based upon an architecture that can be translated easily into multi-chip AER networks. This motivates our description of an expandable DSP/FPGA system that implements similar processing. Finally, we describe the results of recent work in interpreting the outputs of populations of neurons tuned to different disparities. Given a population of neurons tuned to a limited range of disparities, we examine the problem of determining whether or not the input disparity is inside or outside the range of disparities represented in the population. We find that some intuitively appealing measures are actually poor indicators, and suggest that normalization provides a robust way to determine this.

source:Lectures/eebert/BuildingInterpreting.pdf
Auditory Models, Timbre, Top-Down Processing (and why Yahoo! Cares) 

Malcolm Slaney (Yahoo! Research and Stanford CCRMA)

I would like to describe several interesting aspects of auditory perception, setting in motion the discussions I hope we have in the auditory group for the rest of the meeting. I'll talk about temporal processing, timbre perception, top-down processing, and show work that has been done here in Telluride over the years.

source:Lectures/malcolm/AuditoryLecture.pdf
Three Topics in Neural Dynamics 

Jonathan Tapson (University of Cape Town)

This presentation focuses on neural mechanisms for detection of correlation; theory and implementation of the cochlea amplifier; and a brief review of some opinions in event-based control. The topics have been chosen because they connect neuromorphic engineering with mainstream subjects such as statistical signal processing, nonlinear dynamics, and classical control theory. Applications of neuromorphic circuits in fields such as GPS location and sonar ranging are presented.

source:Lectures/jtapson/Three_Topics_in_Neural_Dynamics.pdf
Nic Schraudolph: Something Old, Something New… 

1. Floating-Point Bit-Twiddling for Fun and Profit (PDF)

I will describe extremely fast approximations to exponential, logarithm, power, and logistic functions that I developed years ago but that are now of interest for programming of microcontrollers, GPUs, and other embedded systems.

2. Gradient Methods for Unconstrained Optimization (PDF)

Data-driven, gradient-based optimisation is the engine under the hood of most machine learning techniques. This is a brief review of conventional gradient-based optimization methods, as necessary background for:

3. Stochastic Quasi-Newton Methods for Online Learning (ShockWave Flash) (QuickTime Movie)

By scaling sublinearly with the amount of data, stochastic gradient methods hold unique promise for machine learning and adaptive filtering of large amounts of data. My group has recently developed the first stochastic variants of BFGS and LBFGS, the most popular quasi-Newton methods for nonlinear optimization. I will discuss the modifications necessary for BFGS and LBFGS to work with stochastic approximation of gradients, and report experimental results on convex and non-convex benchmark problems.
John Reynolds: Mapping the microcircuitry of attention 

Authors: Jude F. Mitchell, Kristy A. Sundberg, John H. Reynolds

Cortical neurons differ from one another in important ways, including their neurochemical properties, patterns of connectivity, laminar distribution, gene expression patterns and developmental origin. Previous studies of attention have not sought to distinguish among different classes of neurons. We therefore know almost nothing about the complex circuitry that transforms attentional feedback signals into improved visual processing. Studies in the slice and in anesthetized animals find that parvalbumin expressing GABA-ergic interneurons with the morphologies of basket and chandelier cells have short duration action potentials, whereas most excitatory cell classes have longer duration action potentials, a difference that is due to expression of different classes of sodium and potassium channels. We thus examined differences in attentional modulation across visual area V4 neurons classified on the basis of action potential width. The distribution of action potential widths in our sample of neurons was clearly bimodal. Broad spiking neurons made up the majority of our sample and exhibited markedly lower levels of spontaneous activity and weaker stimulus-evoked responses than narrow spiking neurons. Narrow spiking neurons showed a median increase in firing rate that was substantially larger than the increase that was observed among broad spiking neurons. Attention also reduced response variability, as measured by the Fano factor. This reduction was significant in both types of neurons, but was significantly larger among narrow than broad spiking neurons. This is the first study of attention to distinguish among different neuron types, and our findings lead to the surprising conclusion that attention has a more pronounced influence on local inhibitory interneurons than on pyramidal neurons.
Paul Merolla: Spike communication using address-events 

Analog VLSI offers an efficient way to model neural circuits, but it does not solve the problem of how to connect neurons together --- particularly when they reside on different chips. In this talk, I will focus on the address-event (AE) link, an event-driven asynchronous communication method for modeling axon projections. After describing the nuts and bolts of the AE link, I will introduce how we can connect multiple links together to create large-scale multichip neuromorphic systems.
Patrick Kanold: Vision Tutorial 

source:Lectures/patrickk/vision_tutorial.pdf
Patrick Kanold: How to wire up the brain? 

Neuronal circuits are not predetermined but develop during the prenatal and postnatal period in interplay with the environment. I will highlight some of the processes that are required for establishing the proper functional connectivity of the mammalian brain and that allow the brain to adjust to its environment.

In particular during the prenatal period patterned spontaneous activity is present. In the visual system this activity is present as activity "waves' in the retina and these waves are needed segregation for the segregation of retinal axons into eye specific lamina in the visual thalamus. Specialized learning rules are present in the developing thalamus that interpret the retinal activity patterns and strengthen appropriate connections while weakening inappropriate ones.

During postnatal development neuronal circuits are shaped by sensory experience during "critical periods" of development. After the critical period only limited remodeling is possible. Thus, when peripheral sensory processingi is impaired, cortical circuits can become irreversibly miswired.

The young brain is structurally different from the adult brain and contains additional circuits that are formed by subplate neurons. These neurons are among the earliest born cortical neurons, reside in the white matter and disappear during development. After the critical period ends – when subplate neurons are no longer present - only limited plasticity is present. It is likely that these neurons participate in types of synaptic plasticity that occur only during the critical period.

My work shows that subplate neurons are required for the functional maturation of the cortical columnar organization, the development of intracortical inhibitory circuits and the outcome of plasticity during the critical period. These results suggest that subplate neurons act like a "teacher" helping thalamic neurons to make strong and precise connections to their cortical target neurons. By controlling the balance of excitation and inhibition, subplate neurons influence the correlations between thalamic and cortical activity and thus the amount of cortical activity driven by sensory inputs.

Therefore, I demonstrate how the developing brain complex and dynamically changing circuit which is driven initially by self-generated and later by sensory-evoked activity patterns. I also demonstrate that subplate neurons play a key role controlling early development and plasticity during the critical period by regulating the flow of neuronal activity from thalamus to cortex.
Avis Cohen 
Locomotion: its biological control - onto robots 

The control and generation of the motor pattern underlying locomotion originates in the spinal cord. This lecture presents the evidence for this statement and the organization of the spinal cord that results from it. The interaction with sensory feedback is also described especially in the context of two successful robot implementations of walking: a pair of harnessed biped legs, and an autonomous (tethered) quadrupedal dog (Tekken).

source:Lectures/avis/Tellluride_CPG2007.pdf
Overview of Motor Systems and Motor Control 

In this lecture, we begin with the properties of muscles and and their activators, motoneurons. These are placed in an anatomical context of the spinal cord. The interactions with the interneurons of the spinal cord, and the brain are also described, especially from the motor cortex.

source:Lectures/avis/motor_lecture-07.ppt.pdf
Sue Denham: Multi-stability in Auditory Perceptual Organisation 

The experimental paradigm of auditory two-tone streaming has been used extensively to investigate the processes underlying the formation of sequential associations in auditory scene analysis. However, the classical view of auditory streaming in which it is assumed that the default state is one in which all sounds are integrated into one perceptual object, and that there is then a gradual accumulation of evidence in favour of a segregation of the sounds into two perceptual objects, depending on the stimulus parameters, has been called into question by the finding of bistable switching between different perceptual states. I will present new results from our experiments showing that bistability in auditory is wide-spread; present for all conditions tested and for all subjects. This data shows that many characteristics of auditory bistability appear to be very similar to those found in vision and suggests that some generic mechanisms may underlie both.

source:Lectures/suedenham/Denham.pdf
Richard F. Lyon: Cochlea Modeling Retrospective 

There's a long history of cochlea modeling that people need to be aware of, to help design, optimize, and evaluate neuromorphic hearing systems. In particular, it's important to understand: the notions of time-frequency and time-scale separation and the classes of filters that these notions imply; the large-scale AGC and "essential" nonlinearities that compress the wide dynamic range of sound into a small representation range; the indirect relationship of transfer functions to turning curves; the relative properties of cascade and parallel filterbanks; the need for higher-order poles to get realistic transfer functions; and why and how to capture temporal structure for subsequent processing.

source:Lectures/dicklyon/Lyon_Telluride_Slides.pdf
Chuck Higgins: Reverse-Engineering the Fly 

An Engineer's Approach to the Fly Visual System

In this talk, I motivate why insects are excellent organisms for the study of the neuronal basis of behavior, and describe computational neuroethology experiments in my lab focusing on visual navigation. I provide background on the fly eye and the neurons of its visual system. My talk focuses on two specific projects, one in which behavioral experiments on honeybees have led to a neuronal model of visual navigation, and a second in which the neurons and muscles of a hawkmoth are used to control the motion of an autonomous robot.

source:Lectures/higgins/Higgins_talk.pdf
Shih-Chii Liu: Visual Instruction of the Auditory Spatial Map 

Sound localisation experiments in the barn owl show that the plasticity in auditory spatial maps is guided by visual inputs. We present a model of auditory map formation in the midbrain where plasticity is implemented using STDP. Results show that visual shifts produce equivalent auditory spatial shifts.
Elisabetta Chicca: Cooperation and competition in VLSI networks of spiking neurons 

Abstract to come...

source:Lectures/chicca/Chicca.pdf
Dana Ballard: On the Role of Negative Numbers in Cortical Circuits 

Muller's Law of specific nerve energies introduced the idea that specific nerves transmit information about specific sensory features. This concept has been refined by the notion of `labeled lines,' specific cells that capture specific features of a sensory or motor stimulus, such as Hubel and Weisel's opponent color cells. A further refinement is possible when the feature represented has a range of values distributed about a mean value. In this case such features can be visualized as coding a number that has positive and negative components, where the positive and negative parts are coded with separate nerve cells. We show that special care must be taken in using signed line codings in cortical circuits , particularly feedback circuits. We illustrate the issues using a circuit that learns simple cell encodings from LGN input. This circuit uses a phase-coding of spikes and makes several predictions as to cortical micro-circuitry organization.
Mounya Elhilali: A cocktail party with a cortical twist: A neural and computational view on perceptual sound organization 

The perceptual organization of sounds in the environment into coherent objects is a feat constantly facing the auditory system. It manifests itself in the everyday challenge to humans and animals alike to parse complex acoustic information arising from multiple sound sources into separate auditory streams. While seemingly effortless, uncovering the neural mechanisms and computational principles underlying this remarkable ability remain a challenge facing both the biological and mathematical communities. In this talk, I discuss how this perceptual ability of the auditory system may emerge as a consequence of a multi-scale spectro-temporal analysis of sound in the auditory cortex, which is thought to play a role in the perceptual ordering of acoustic events. In addition, I present recent findings of adaptive neuronal responses in the auditory cortex, which are likely to play a key role in adapting the neural representation to reflect both the sensory content and the changing behavioral context of complex acoustic scenes. Guided by these physiological results, I shall present a computational approach to dynamic segregation of auditory streams, based on unsupervised clustering and the statistical theory of Kalman filtering.
Michael Stryker: Creating Maps in the Brain 

The visual cortex is organized into retinotopic maps that preserve an orderly representation of the visual world, achieved by topographically precise inputs from the lateral geniculate nucleus. We first showed that geniculocortical mapping is imprecise when the waves of spontaneous activity in the retina during the first postnatal week are disrupted genetically. This anatomical mapping defect is present by postnatal day 8 and has functional consequences as revealed by optical imaging and microelectrode recording in adults. Pharmacological disruption of these retinal waves during the first week phenocopies the mapping defect, confirming both the site and the timing of the disruption in neural activity responsible for the defect. Analysis shows that the geniculocortical miswiring is not a trivial or necessary consequence of the retinogeniculate defect. This demonstrated that disrupting early spontaneous activity in the eye alters thalamic connections to the cortex.

What are the molecular cues that guide development of visual cortical maps? Ephrin-As and their receptors, EphAs?, are expressed in developing cortex where they may act to organize thalamic inputs. We mapped the visual cortex in mice deficient for ephrin-A2, -A3, and -A5 functionally, using intrinsic signal optical imaging and microelectrode recording, and structurally, by anatomical tracing of thalamocortical projections, and find that the visual cortex is shifted medially, rotated, compressed, and the internal organization of its map is degraded. Expressing ephrin-A5 ectopically by in utero electroporation in the lateral cortex shifts the map of the visual cortex medially, and expression within the visual cortex disrupts the internal organization of the map. These findings indicated that interactions between gradients of EphA/ephrin-A in the cortex guide map formation but that factors other than redundant ephrin-As are responsible for the remnant map.

The topographic map depends on both neural activity and ephrin-A nmolecular guidance cues. To determine whether these mechanisms act in parallel or in series, we studied mice deficient in both ephrin-As and retinal waves. We found that the functional and anatomical cortical maps in these mice are nearly abolished along the naso-temporal (azimuth) axis of the visual space, a disruption much more severe than is produced by removing either signal alone. These results demonstrate that ephrin-As and structured neuronal activity are two independent pathways mediating map formation in the visual cortex and that together, they account almost completely for the formation of the azimuth map. Strikingly, despite the near abolition of the azimuth map, the elevation map is qualitatively normal, indicating that the two Cartesian axes of the cortical map are organized by distinct mechanisms.

source:Lectures/stryker/Telluride-Stryker-2007.pdf
Jimmy Abbas: Neuromorphic Design of Smart Prosthetic and Therapeutic Systems 

Neuromophic approaches to prosthetic control systems may help to endow the next generation of prostheses with enhanced functionality and facilitate more rapid and complete integration with residual control system. Similarly, therapeutic devices that are based on neuromorphic principles may enable more effective and efficient therapy both in the clinic and home environments. This talk will present examples of prosthetic and therapeutic technology currently in development and discuss opportunities and challenges for utilizing neuromorphic engineering approaches in medical rehabilitation.

source:Lectures/jabbas/Abbas_-_INE_Telluride_-_070714.pdf
Kwabena Boahen: How to outperform a supercomputer with neuromorphic hardware 

The digital technique used to simulate neural activity has not changed since Hodgkin and Huxley pioneered ion-channel modeling in the 1950s. Since then, progress has come incrementally, from the doubling in computer performance every eighteen months (Moore’s Law), plateauing in recent years, putting real-time cortex-scale simulations outside the realm of the fastest supercomputers for the foreseeable future. Fortuitously, the analog technique investigated by neuromorphic engineers over the past two decades has now matured, with the recently developed ability to program various types of ion-channels as well as arbitrary patterns of synaptic connections. This radical technique is poised to deliver beyond-Blue-Gene performance on a Beowulf-cluster budget.

source:Lectures/kwabena/Neurogrid_Tel07.pdf
Jochen Triesch: Synergies Between Intrinsic and Synaptic Plasticity Mechanisms 

Different forms of neural plasticity shape cortical representations. Intrinsic plasticity refers to a neuron's ability to change its nonlinear properties. I propose a model of intrinsic plasticity for a continuous activation model neuron based on information theory. I show how intrinsic and synaptic plasticity mechanisms can synergistically interact to allow the neuron to discover heavy-tailed directions in the input space. This capability can be used for ICA-like learning.

source:Lectures/triesch/Telluride2007-short.pdf
Giacomo Indiveri: Spike-based learning and sensory processing in silicon 

As the number of VLSI implementations of spikebased neural networks is steadily increasing, and the development of spike-based multi-chip systems is becoming more popular it is important to design spike-based learning algorithms and circuits, compatible with existing solutions, that endow these systems with adaptation and classi?cation capabilities. I describe a spike-based learning algorithm that is highly effective in classifying complex patterns in semi-supervised fashion, and present neuromorphic circuits that support its VLSI implementation. I also show what is required to construct complex (learning) neuromorphic systems and argue that we as a community need many more resources and investments before claiming that we can solve real-world problems in a robust and reliable way.

source:Lectures/giacomo/telluride07.pdf
Mikko A. Uusitalo: Nanotechnology and neuromorphic engineering 

source:Lectures/mikkou/Nokia_Uusitalo_Telluride07.pdf
Ralph Etienne-Cummings: Generation and Control of Spinal Locomotion Signals and Their Application to Biomorphic Robots 

Abstract here...

source:Lectures/retienne0/Tell07_REC.pdf


Slides only: 

    * source:Lectures/shinn/Shinn.pdf
    * source:Lectures/shinn/Shinn.2.pdf 

    * source:Lectures/ijspeert/Telluride_July2007.pdf 


Do not attach PDFs here, upload them to subversion (Lectures) or send them to the SysAdmin.
Download in other formats:

    * Plain Text

Trac Powered

Powered by Trac 0.10.5dev
By Edgewall Software.

Telluride Workshop 2007