Detection of spike patterns using pattern filtering, with applications to sleep replay in birdsong

doi:10.1016/S0925-2312(02)00746-4

Neurocomputing

Volumes 52–54, June 2003, Pages 19-24

https://doi.org/10.1016/S0925-2312(02)00746-4 Get rights and content

Abstract

Consolidation of memories during sleep may involve neuronal replay of daytime activity patterns. A precise replay phenomenon was observed in the song system nucleus RA. We developed statistical pattern filtering to investigate replay in HVc, where spiking activity is more variable. Long spike sequences during sleep were convolved with filters constructed from exemplar auditory spike sequences. Examining peaks in the filter responses, we found many examples of HVc spike trains with temporal patterns similar to the exemplars, confirming a replay phenomenon for HVc. Statistical pattern filtering has general utility for finding spike patterns when a referent pattern is known.

Introduction

In the zebra finch forebrain nucleus robustus archistriatalis (RA), neurons generate sequences of spike bursts during spontaneous sleep activity that exhibit similar temporal patterns as the pre-motor activity during singing and the auditory responses to song playback in the sleeping animal [2], [3]. This sleep replay phenomenon is hypothesized to play an important role in learning and memory consolidation of the birdsong system [2], [7], which gives rise to the question of whether and how other parts of the song system participate in replay during sleep. The forebrain nucleus HVc is a sensorimotor center of the birdsong system. It projects to RA, and is thought to play an important role in vocal learning, auditory input integration, and higher level motor command. State-dependent auditory responses have also been observed for the HVc [9], [10]. We thus investigated whether the replay phenomenon also occurs in spontaneous sleep activity of HVc.

Instead of searching for individual replayed spike bursts [2], our goal was to find burst sequences that have similar temporal patterns as pre-motor activity or auditory responses [6], [8]. This is challenging for two reasons. First, unlike pre-motor or auditory data, for sleep spontaneous neuronal data, there are no reliable observables to mark the occurrence of neuronal events of interest. Second, the neuronal activity in HVc has far more variability than RA activity.

Section snippets

Pattern filtering

We developed a statistical methodology using linear filtering to address the above difficulty. In general, let S be an exemplar train of spikes registered at s₁<s₂<⋯<s_p, s_k∈[0,σ]. Given a small number ε>0, fix two functions K(x) and B(x), with K(x)≠0 only for x∈(−ε,ε) and B(x)⩽0 for all x. The functions will be referred to as the “time window function” for a spike and the “background function”, respectively, and ε the “window size”. Define F on [0,σ] by $F(x)= max_{1⩽k⩽p} K(x−s_{k}) if x∈J_{S}, B(x) otherwise,$

The underlying statistical model

Pattern filtering is a likelihood ratio test, as opposed to likelihood test in the unitary event analysis [4], [5]. To illustrate, consider the simplest filter constructed from the time window function K(x)≡α>0, x∈(−ε,ε), and the background function B(x)≡−β⩽0. Then (3) leads to $R(x)=#{t_{i} ∈T∩[x,x+σ] : t_{i} −x∈J_{S}}×α−#{t_{i} ∈T∩[x,x+σ] : t_{i} −x∉J_{S}}×β$ . Because J_S⊂[0,σ], the formula is simplified into $R(x)=#{t_{i} ∈T : t_{i} −x∈J_{S}}×α−#{t_{i} ∈T : t_{i} −x∉J_{S}}×β.$

Detection using this filter is derived from three statistical

Results

We implemented pattern filtering for HVc single unit data recorded during sleep. For each unit, the spike sequences in response to the bird's own song (‘BOS’) broadcast to the sleeping bird in multiple trials were selected as templates. Then, in the subsequent spontaneous activity of the unit when no acoustical stimulus was presented, segments with similar temporal patterns as the templates were searched. For the neuronal activity in HVc, several types of variability have to be taken into

Zhiyi Chi entered Peking University, Beijing, China in 1989. He entered Brown University, Providence, Rhode Island in 1993. He received an Sc.M.from the Department of Mathematics the next year. In 1998, he graduated from the Division of Applied Mathematics with a Ph.D. for work on ‘Probability Models for Complex Systems’ under the supervision of Stuart Geman. Since then he has been Assistant Professor in the Department of Statistics at the University of Chicago. He is currently also a member of

References (10)

K. Louie et al.
Temporally structured replay of awake hippocampal ensemble activity during rapid eye movement sleep
Neuron
(2001)
D. Margoliash
Do sleeping birds sing? Population coding and learning in the bird song system
Prog. Brain Res.
(2001)
M. Abeles et al.
Detecting spatio-temporal firing patterns among simultaneously recorded single neurons
J. Neurophysiol.
(1998)
A.S. Dave et al.
Song replay during sleep and computational rules for sensorimotor vocal learning
Science
(2000)
A.S. Dave et al.
Behavioral state modulation of auditory activity in a vocal motor system
Science
(1998)

There are more references available in the full text version of this article.

Cited by (3)

Set and setting: How behavioral state regulates sensory function and plasticity
2013, Neurobiology of Learning and Memory
Citation Excerpt :
The sleep-associated, “auditory-like” song responses in RA are similar to those normally seen in wakefulness in auditory-responsive neurons which project to RA from the HVC (high vocal center) nucleus (Dave & Margoliash, 2000; Dave et al., 1998). During sleep, HVC neurons burst in a temporal pattern which mimics the pattern associated with the sound of the bird’s own song (Chi, Rauske, & Margoliash, 2003). The gating of HVC auditory input to RA during wakefulness is mediated by noradrenergic and cholinergic signaling within HVC; local agonist application within HVC, or basal forebrain stimulation leads to suppression of auditory responses in RA, but not HVC (Dave et al., 1998; Rauske, Shea, & Margoliash, 2003).
Recently developed neuroimaging and electrophysiological techniques are allowing us to answer fundamental questions about how behavioral states regulate our perception of the external environment. Studies using these techniques have yielded surprising insights into how sensory processing is affected at the earliest stages by attention and motivation, and how new sensory information received during wakefulness (e.g., during learning) continues to affect sensory brain circuits (leading to plastic changes) during subsequent sleep. This review aims to describe how brain states affect sensory response properties among neurons in primary and secondary sensory cortices, and how this relates to psychophysical detection thresholds and performance on sensory discrimination tasks. This is not intended to serve as a comprehensive overview of all brain states, or all sensory systems, but instead as an illustrative description of how three specific state variables (attention, motivation, and vigilance [i.e., sleep vs. wakefulness]) affect sensory systems in which they have been best studied.
Continuous speech recognition with sparse coding
2009, Computer Speech and Language
Sparse coding is an efficient way of coding information. In a sparse code most of the code elements are zero; very few are active. Sparse codes are intended to correspond to the spike trains with which biological neurons communicate. In this article, we show how sparse codes can be used to do continuous speech recognition. We use the TIDIGITS dataset to illustrate the process. First a waveform is transformed into a spectrogram, and a sparse code for the spectrogram is found by means of a linear generative model. The spike train is classified by making use of a spike train model and dynamic programming. It is computationally expensive to find a sparse code. We use an iterative subset selection algorithm with quadratic programming for this process. This algorithm finds a sparse code in reasonable time if the input is limited to a fairly coarse spectral resolution. At this resolution, our system achieves a word error rate of 19%, whereas a system based on Hidden Markov Models achieves a word error rate of 15% at the same resolution.
From gestalt theory to image analysis: A probabilistic approach
2008, Interdisciplinary Applied Mathematics

Peter Rauske studied physics at the University of Chicago, graduating in 1994. He is currently studying the neurobiology of birdsong production and learning using chronic recording techniques with Daniel Margoliash's group at the University of Chicago, where he is pursuing a Ph.D. in the Department of Organismal Biology and Anatomy.

Daniel Margoliash studied biology at the California Institute of Technology. After graduating in 1975 he entered the Bioinformation Systems program in Engineering Science at Caltech. Eventually joining the laboratory of Masakazu Konishi, he studied auditory responses in the birdsong system. He received an M.Sc. in 1981, and in 1983 he received his Ph.D. for work on ‘Songbirds, Grandmothers, and Templates: A Neuroethological Approach’. After a brief postdoctoral fellowship with Nobuo Suga at Washington University, St. Louis, he joined the faculty of the then Department of Anatomy at the University of Chicago in 1986. He is currently Professor in the Departments of Organismal Biology and Anatomy & Psychology, and the Committees on Neurobiology & Computational Neuroscience at the University of Chicago.

View full text

Detection of spike patterns using pattern filtering, with applications to sleep replay in birdsong

Abstract

Introduction

Section snippets

Pattern filtering

The underlying statistical model

Results

Neuron

Prog. Brain Res.

Detecting spatio-temporal firing patterns among simultaneously recorded single neurons

J. Neurophysiol.

Song replay during sleep and computational rules for sensorimotor vocal learning

Science

Behavioral state modulation of auditory activity in a vocal motor system

Science