Cosyne 2009 Workshops
March 2-3, 2009
Snow Bird, Utah
Workshop Title
Stochastic methods for the analyses of spike train and field potential data
Organizer(s)
David Nguyen, MIT
Uri Eden, BU
Download Slides
Abstract
Stochastic methods have proven to be versatile and appropriate for identifying structured activity in multi-unit spike train and/or field potential recordings. In our attempts to understand how the brain represents and processes information, the divide and conquer strategy has been most prevalent. A consequence of specialization by neural systems is the development and application of important analyses methods may occur independently. It is the goal of this workshop to cover standard and cutting edge methods used in the study of systems such as vision, auditory, motor, hippocampus, and disease with the hope that identified common needs may direct future collaborations and research, while differences in approach may spark healthy discussions that would lead to a transfer of expertise across specializations.
Speakers
8:00 - 8:55 Uri Eden (BU)
Using point process models to encode and decode spike train data
9:00 - 9:55 Rob Haslinger (MIT)
Including Local Field Potentials in Parametric Models of Neuronal Spiking
10:00 - 10:55 Gordon Pipa (Max-Planck)
Detecting Synchronized spiking activity: Theoretical and practical issues elements
(Break)
4:30 - 5:25 Maneesh Sahani (UCL)
Gaussian process methods for single-trial analysis of dynamical data
5:30 - 6:25 Terry Sanger (Stanford)
Discrete neurons in a continuous world; Understanding dynamics and control by assemblies of stochastic quantized
6:30 - 7:25 Carina Curto (Rutgers)
Using cross-validation to test models of neuronal spiking activity
Speaker Abstracts
Uri Eden
Using point process models to encode and decode spike train data
Although it is well known that neurons receive, process and transmit signals via sequences of sudden stereotyped electrical events, many analyses of neural data ignore the highly localized nature of spikes. The theory of point processes offers a unified, principled approach to modeling and estimating the firing properties of spiking neural systems, and assessing goodness-of-fit between a neural model and observed spike train data.
We develop a state space estimation framework to track the evolution of dynamic signals using spike train observations from neural ensembles. This allows us to derive a toolbox of estimation algorithms and adaptive filters to address questions of static and dynamic encoding and decoding. In our analysis of these filtering algorithms, we draw analogies to well-studied linear estimation algorithms for continuous valued processes, such as the Kalman filter and its discrete and continuous time extensions.
These methods will be illustrated in the context of the analysis of place field activity in the rodent hippocampus. Place cells, which tend to fire preferentially when the animal is in specific locations, have been implicated in cognitive tasks such as navigation and decision making. Using simple point process models, we are able to accurately characterize the localized spiking activity of these neurons as a function of the animal's position in its environment, track plasticity in their firing properties, and reconstruct the animal's movements from the spiking of a hippocampal population.
Rob Haslinger
Including Local Field Potentials in Parametric Models of Neuronal Spiking
A neuron's spiking depends upon many factors including not only external stimuli and the previous spiking history of the neuron in question, but also the activity of the neural network in which the neuron is embedded. Recording the activity of large populations of neurons is technically difficult and for this reason the Local Field Potential (LFP) is often used as a surrogate measure of network activity. Deducing the extent to which, and manner in which, the LFP predicts a neuron's spiking is non-trivial. The spatially non-local nature of the LFP makes biophysical interpretations problematic. Phenomenological approaches, employing parametric regression Generalized Linear Models) can be used to determine whether a given LFP feature (such as spectral power or phase) is predictive of LFP spiking. We discuss the technical details for doing so, and also some new approaches for determining which LFP features to use as covariates in the regression. These ideas are applied to neurons and LFPs recorded in awake V1 macaque cortex.
Gordon Pipa
Detecting Synchronized spiking activity: Theoretical and practical issues.
It is commonly held that neurons encode information by modulations of their discharge rate. A complementary hypothesis is, that information is also encoded in the precise relation between the discharges of spatially distributed neurons. These complementary views are addressed in the literature as the rate coding and the temporal coding hypothesis. Multiple methods have been developed to detect temporal relations between spiking events and to investigate whether these relations that are forming a spike pattern are correlated with stimulus configurations, behavior, or particular states of neuronal systems. The methods differ in the definitions of the spike patterns, the techniques to detect these patterns, and the approaches to analyze the resulting data (descriptive, statistical hypothesis testing, maximum likelihood, and Bayesian approaches). Even though the temporal coding hypothesis formulates precisely what constitutes a spike pattern, it turns out to be a non-trivial problem to design a method that detects the existence of such pattern, and investigates their information content, without being confounded by other properties of the data.
In this workshop we are going to present new a non-parametric and computationally-efficient method named NeuroXidence (see www.NeuroXidence.com) that detects coordinated firing within a group of two or more neurons and tests whether the observed level of coordinated firing is significantly different from that expected by chance. NeuroXidence (1) considers the full auto-structure of the data, including the changes in the rate responses and the history dependencies in the spiking activity. We demonstrate that NeuroXidence can identify epochs with significant spike synchronisation even if these coincide with strong and fast rate modulations. We also show, that the method accounts for trial-by-trial variability in the rate responses and their latencies, and that it can be applied to short data windows lasting only tens of milliseconds. Based on simulated data we compare the performance of NeuroXidence with the UE-method and the cross-correlation analysis.
In this talk, I will cover theoretical background, practical guidelines, and hands on demonstrations of the tool NeuroXidence. The Matlab Toolbox NeuroXidence (see www.NeuroXidence.com) will be provided
Maneesh Sahani
Gaussian process methods for single-trial analysis of dynamical data
Traditional methods for analysing spike data, which rely on averaging across multiple repeated trials, break down when neural activity is likely to fluctuate from trial to trial. Unfortunately, the sparseness of spike data makes single-trial analysis difficult. Sensible priors, either implicit or explicit, on the underlying activity become paramount. The simplest prior---that of constant intensity during the trial---is rarely suitable. We have developed a family of methods that use the infinitely-parameterised Gaussian process prior to estimate and visualise neural activity. These methods may be applied to single neurons or to multineuron recordings, and may be used with point-process likelihoods (both conditionally Poisson and with conditionally self-exciting) or likelihoods on spike counts. I will discuss the inference and hyperparameter optimisation issues within this family.
Joint work with Byron Yu, John Cunningham and Krishna Shenoy.
Terry Sanger
Discrete neurons in a continuous world; Understanding dynamics and control by assemblies of stochastic quantized dynamic systems
Neural coding is often described as "noisy" in the sense that computation based on neural data requires statistical estimation procedures. But when neural systems must perform rapid sensory estimation or dynamic control, then there is no time to perform statistical estimation. Nevertheless, estimation and control are possible on natural time-scales because under certain circumstances the expected behavior of populations of neurons accurately reflects desired computations at very short latency. The population average can be constructed to be an unbiased estimator of a desired function. The variance of any sample estimate is determined by the number of neurons and the time available for observation, so these results provide a direct explanation for psychophysical results including the speed-accuracy tradeoff for movement. These results provide a new framework for understanding computation in large populations of spiking neurons. An important and possibly counter-intuitive result is that "noisy" neurons are in some cases more efficient for transmission of high-dimensional information than deterministic high-resolution measurements.
Carina Curto
Using cross-validation to test models of neuronal spiking activity
Neural population recordings produce vast amounts of high-dimensional data, which poses a challenge for model-based statistical methods. Cross-validation provides a simple and robust framework with which to analyze this kind of data, because errors in model specification (including the addition of extraneous parameters) can only lead to worse, not better performance. We describe two studies in which cross-validation was used to fit and test models of spiking activity in electrophysiological recordings. In the first study, models for hippocampal place cell firing are compared. In the second study, we fit a low-dimensional dynamical systems model to population activity in primary auditory cortex, and verify that the dynamics are predictive of click-evoked sensory responses.
