Cosyne 2009 Workshops
March 2-3, 2009
Snow Bird, Utah
Workshop Title
Neural variability and movement variability
Organizer
Mark Churchland, Stanford
Abstract
In 1990, Larry Bird made 71 consecutive free throws. While a remarkable feat, one wonders why he missed the 72nd? Why could he not simply do what he had done the last 71 times? We generally take for granted that repeated ‘trials’ of the same movement will have variable results, but why is this so?
A variety of explanations have been proposed to explain across-trial behavioral variability. First, variability could actually aid learning by probing the local parameter space. Second, it is known that the optimal control strategy is to allow variability in non-critical task dimensions. Reducing such variability would lead to higher error for the critical dimensions. Third there are likely many solutions (e.g., many quasi-optimal controllers) to any given control problem. When different solutions are found on different trials, variability (most benign, some not) would result. Finally, variability may simply be endemic to large recurrent networks with non-linear units. Such variability might be reduced during movement, but never completely quashed. It is possible that all of the above explanations are partially correct, or that none are.
Investigating such issues requires tracking behavioral variability to its neural basis. The goal of this workshop is to describe recent work to that end. We will focus on the following questions. First, as most neural recordings are remarkably variable, when does this variability translate to behavioral variability, and when doesn’t it? Is the noisy spiking of individual neurons an important source of variability? Or is the critical factor correlated variability in the underlying rate? If the latter, how can we make measurements that isolate that variability? When can we produce sufficiently-concrete models to generate testable predictions regarding variability, at both the behavioral and neural levels?
Ultimately we wish to make progress on the following question: how should we view, analyze and interpret our variable neural recordings, if we are to explain behavioral variability?
Speakers
Morning
8:00 - 8:10 Brief introduction
8:10 - 8:35 Stephen Scott (Queens University, Canada)
8:35 - 8:45 Discussion
8:45 - 9:10 Stephen Lisberger (UCSF)
9:10 - 9:20 Discussion
9:20 - 9:40 Coffee break
9:40 - 10:05 Michael Brainard (UCSF)
10:05 - 10:15 Discussion
10:15 - 10:40 Daniel Wolpert (Cambridge)
10:40 - 11:00 Extended discussion
Afternoon
4:30 - 4:55 Peter Latham (Gatsby, London)
4:55 - 5:05 Discussion
5:05 - 5:30 Martin Nawrot (Free University Berlin)
5:30 - 5:40 Discussion
5:40 - 6:05 Coffee break
6:05 - 6:25 David Sussillo (Columbia University)
6:25 - 6:35 Discussion
6:35 - 7:00 Mark Churchland (Stanford University)
7:00 - 7:30 Extended Discussion
Speaker Abstracts
Stephen Scott (Queens University, Canada) Interpreting motor behaviour with stochastic optimal feedback control
My presentation will provide a very general overview of the principles of optimal feedback control through simple didactic examples. I will draw from studies on a range of animal and human studies to illustrate how aspects of optimal feedback control can explain patterns of variability, synergies and redundancy. Finally, I will describe some recent experiments we have been developing that illustrate how the motor system can take advantage of motor redundancy between the two limbs during bimanual motor skills.
Stephen Lisberger (UCSF) The relationship between neural and motor variation
The brain is intrinsically variable, as is its output in the form of behavior. We have been using smooth pursuit eye movements to ask about the relationship between neural and behavioral variation, and to divide neural variation into components that are signal versus noise. We have found a surprisingly large correlation between the trial-by-trial variation in pursuit eye velocity and the concomitant variation in neural responses in both the floccular complex of the cerebellum and the smooth eye movement region of the frontal eye fields in the cerebral cortex (FEFSEM). The strong “neuron-behavior” correlations imply that some of the neural variation is signal that is shared across neurons and drives variation in behavior, while the rest of the neural variation in noise that is reduced downstream from each site. Recordings from both the FEFSEM and the flocular complex imply that essentially no noise is added in downstream processing, implying that all the variation in the behavior arises at earlier stages, perhaps in the sensory processing for pursuit.
Michael Brainard (UCSF) Sources and function of variation in adult birdsong
Adult birdsong is an example of a well-learned and highly stereotyped motor skill. Nevertheless, for adult song, as for all motor skills, there is subtle residual variation in performance from one iteration to the next. Such variation is often construed as biological noise, perhaps arising from the motor periphery, that is below threshold for behavioral relevance and not subject to central neural control. I will describe experiments that address an alternative hypothesis, that such behavioral variation, rather than meaningless noise, represents subtle, but active experimentation by the nervous system to optimize and maintain motor performance. I will first describe neurophysiological experiments that indicate a central neural origin for a component of observed variation in features of song such as the pitch of individual syllables. These experiments indicate that excess variation is actively introduced into song premotor circuitry from extrinsic sources, including avian basal ganglia structures know to be important for learning. I will then show that the subtle variation present in song can be used by the nervous system to enable rapid adaptive plasticity of 'crystallized', post-critical period song. Collectively, these experiments suggest that for birdsong, and perhaps other well-learned vertebrate skills, subtle behavioral variation may be purposefully introduced by the nervous system to enable continuous monitoring and optimization of performance.
Daniel Wolpert (Cambridge, sub. for Aldo Faisal) How does the nervous system organise behaviour in the face of internal variability & time constraints?
Noise affects all levels of information processing from the interaction of signalling molecules in neurons, the responses of neural circuits, to human limb movements. Noise has emerged as a key ingredient in shaping the structure and function of the nervous system (Faisal et al, 2008, Nature Rev. Neurosci.) and results in uncertainty and variability in both perception and action. Time is a critical factor in countering the effects of noise, either by allowing to average out noise over time or acquire novel information to reduce uncertainty. In most natural situations however time is an essential factor that the brain has to account when faced with timely decisions in a dynamic environment. Here, I present evidence from sensorimotor studies in humans that suggest how the brain treats both processing and movement variability in the face of time constraints. The experimental approach and simple probabilistic framework allow us to predict the behavioural performance in a naturalistic task which uses time as a free parameter (unlike the classic and fairly artificial reaction time task).
Peter Latham (Gatsby, London) Noise, approximate inference, and behavior.
It is well known that neuronal activity in the brain is variable, in the sense that identical sensory stimuli produce different responses. It has been difficult to determine, however, what that variability means. Is it noise, or does it carry important information - about, for example, the internal state of the organism? Here we argue, based on whole-cell, in vivo, recordings, that the brain is highly chaotic and, therefore, most of the observed variability is noise. We also argue that the kind of noise induced by chaotic dynamics (and similar sources, such as stochastic neurotransmitter release) have little effect on behavioral variability; most of the behavioral variability we see is due to the approximations the brain must make to solve real-world problems.
Martin Nawrot (Free University Berlin) The effect of cortical network state evolution on the amount and dynamics of single neuron variability
The output of individual cortical neurons as recorded in the living brain shows a high response variability across experimental repetitions (e.g. Shadlen & Newsome, 1998; Lee et al., 1998). Yet, the cortex is able to process sensory information with an intriguing temporal fidelity and behavioral responses display a high accuracy in space and time. To solve this apparent contradiction we combined various in vitro and in vivo experimental approaches with model-based stochastic analyses of count and interval statistics.
We distinguish (i) trial-by-trial variability of the spike count as quantified by the Fano factor (FF), and (ii) inter-spike interval variability as measured by the coefficient of variation (CV) of the ISI distribution. I will briefly report on the conditions that allow for a bias-free estimate and the method we used to measure the CV under rate-varying conditions. We suggest a joint analysis of both measures and interpret deviations from the renewal prediction FF=CV^2 (Nawrot et al., 2008). A time-resolved analysis of FF and CV^2 allows to access and interpret task-related variability dynamics as observed in behaving animals (e.g. Nawrot et al., 2003; Churchland et al., 2006; Nawrot et al., 2008).
Somatic noise current injection in vitro, and in vivo intracellular recordings under anaesthetized conditions (Exp. by C Boucsein and M Nawrot) allowed us to quantify single neuron output variability under stationary input conditions (Nawrot et al., 2007; Nawrot et al., 2008). Analysis of multiple single unit recordings from the motor cortex of behaving monkeys (Exp. in the lab of A Riehle) allowed us to estimate amount and dynamics of spike train variability in vivo. Observation of large scale brain signals such as the LFP and human epicortical field potentials were used to monitor network activity on global spatial and temporal scales.
Our results imply that the Poisson point process (FF=CV^2=1) is a deficient model for the description of spike train statistics of real cortical neurons which are clearly less variable (FF=CV^2<1) under conditions of balanced excitatory and inhibitory input. Thus, only about half of the observed single neuron variability in vivo (FF>1) can be explained by the stochastic nature of the balanced input and of synaptic transmission. I present a simple model that can explain how the ongoing global network state dynamics can boost cortical variability, and how it influences task-related variability dynamics (Nawrot 2003).
David Sussillo (Columbia University) Generating coherent patterns of activity from chaotic neural networks
Neural circuits typically display a rich set of spontaneous activity patterns, and also exhibit complex activity when responding to a stimulus or generating a motor output. How are these two forms of activity related? We develop a procedure we call FORCE learning for modifying feedback loops either external to or within a model neural network to change chaotic spontaneous activity into a wide variety of desired activity patterns. FORCE learning, which involves error-directed modifications of synaptic strengths, works even though the networks we train are chaotic and, in contrast to earlier proposals, we leave all feedback loops intact and all neurons unconstrained during learning. Using this approach, we construct networks than produce a wide variety of complex output patterns, input-output transformations that require memory, multiple outputs that can be switched by control inputs, and motor patterns matching human motion capture data. Our results reproduce recent data on pre-movement activity in motor and premotor cortex, and suggest that synaptic plasticity may be a more rapid and powerful modulator of network activity than generally appreciated.
Mark Churchland (Stanford University) A pan-cortical stimulus-driven decline in neural variability
Much of systems neuroscience is essentially a systems identification procedure: probing a system, observing its response, and inferring its dynamics. An engineer would wish to analyze every trial’s response. Yet in neuroscience often only the mean across-trial response is considered. The motivation for averaging is understandable, given the noisy spiking statistics of individual neurons. Yet while averaging can combat spiking noise, it presumably also masks real variability in the underlying rate. To recover some of the lost information, we extended current mathematical methods (the Fano factor and factor analysis) to reveal the envelope of across-trial rate variability. In each of seven brain areas spanning the four cortical lobes, across-trial variability was substantial and declined following stimulus onset. Thus, changes in across-trial variability – normally obscured by averaging – are a widespread feature of the brain’s response to stimuli. The consistent decline argues that the relevant neural circuits are stabilized by an input, in agreement with recent computational predictions.
