Discovering frequent episodes in event sequences is an interesting data mining task. In this paper, we argue that this framework is very effective for analyzing multi-neuronal spike train data. Analyzing spike train data is an important problem in neuroscience though there are no data mining approaches reported for this. Motivated by this application, we introduce different temporal constraints on the occurrences of episodes. We present algorithms for discovering frequent episodes under temporal constraints. Through simulations, we show that our method is very effective for analyzing spike train data for unearthing underlying connectivity patterns.
Temporal data mining is concerned with mining of large sequential data sets [8]. Frequent episode discovery, originally proposed in [11], is one of the popular frameworks in temporal data mining. Here the data is viewed as a single long sequence of events and the task is to unearth temporal patterns (called episodes) that occur sufficiently often along that sequence. Examples of such data are alarms in a telecommunication network, fault logs of a manufacturing plant, multi-neuronal spike train recordings, etc.
In this paper we present new algorithms for discovering frequent episodes under some temporal constraints. The motivation for considering such constraints comes from the application that we discuss here, namely, analyzing multineuron spiking data to infer useful information about the underlying microcircuits. Such neuron spike train data can be obtained through techniques such as microelectrode array experiments. Analyzing simultaneously recorded data from a number of neurons is an important and challenging problem. The data consists of spike trains from a number of neurons. Since functionally interconnected neurons tend to fire in certain precise patterns, discovering frequent patterns in such temporal data can help understand the underlying neural circuitry. Here, we argue that the frequent episodes framework is ideally suited for such analysis. However, as we shall see, for this application we need methods to discover frequent episodes where the occurrences of episodes need to satisfy some additional temporal constraints. The currently available methods for frequent episode discovery can not tackle such constraints. In this paper, we present some new algorithms for frequent episode discovery under such temporal constraints.
We explain the problem of analyzing multi-neuronal spike train data in Section II. We then present a brief overview of the frequent episodes framework in Section III. We introduce the notion of temporal constraints on the episode occurrences and explain how one can use methods of serial and parallel episode discovery under temporal constraints, to discover many patterns of interest in the spike train data. The algorithms for discovering frequent episodes under temporal constraints are presented in Section IV. We present some simulation results to illustrate our method of discovering connection patterns in neuronal networks in Section V. Finally, we conclude the paper in Section VI with a discussion.
Over the last couple of decades many new technologies have made it possible to simultaneously record signals from many neurons and hence to study microcircuits in neuronal assemblies. Microelectrode array (MEA) is one such popular recording technology. A typical MEA setup consists of 8 × 8 grid of 64 electrodes with inter-electrode spacing of about 25 microns and can be mounted on a neural culture or brain slice. Other technologies for recording from multiple neurons include imaging of neuronal currents using some specialized dyes. These technologies now allow for gathering of vast amounts of data, especially in neuronal cultures, using which one wishes to study connectivity patterns and microcircuits in neural systems. (See [13], [6] for some recent studies of this kind).
The availability of vast amounts of such data means that developing efficient methods to analyze neuronal spike trains is a challenging task of immediate utility in this area. A recent review by Brown et.al. summarizes the current state of art [4]. Most of the current methods of analysis rely on quantities that can be computed through cross correlations among spike trains (time shifted with respect to one another) to identify interesting patterns in spiking activity [4]. There are also methods that look for specific fixed patterns and assess their statistical significance under a null hypothesis that different spike trains are iid Bernoulli processes [2], [10], [12]. Most such methods can not look for patterns that involve more than 3 or 4 neurons due to the ubiquitous curse of dimensionality. Hence model-free techniques such as data mining can be very useful in unearthing interesting patterns in the spike trains.
The patterns that one is interested in this application can be roughly grouped into what are called Synchrony, Order and Synfire chains. Synchronous firing by a group of neurons is interesting because it can be an efficient way to transmit information [5]. Ordered firing sequences of neurons where times between firing of successive neurons are fairly constant denote a chain of triggering events and unearthing such relations between neurons can thus reveal some microcircuits [1]. Such an ordered chain may be among groups of neurons rather than single neurons. Such a pattern is called a Synfire chain and is believed to be a very important microcircuit [6]. In the next section, we explain how all such patterns can be discovered under the framework of frequent episodes with temporal constraints.
Frequen
This content is AI-processed based on open access ArXiv data.
