When analyzing the neuronal code, neuroscientists usually perform extra-cellular recordings of neuronal responses (spikes). Since the size of the microelectrodes used to perform these recordings is much larger than the size of the cells, responses from multiple neurons are recorded by each micro-electrode. Thus, the obtained response must be classified and evaluated, in order to identify how many neurons were recorded, and to assess which neuron generated each spike. A platform for the mass-classification of neuronal responses is proposed in this chapter, employing data-parallelism for speeding up the classification of neuronal responses. The platform is built in a modular way, supporting multiple web-interfaces, different back-end environments for parallel computing or different algorithms for spike classification. Experimental results on the proposed platform show that even for an unbalanced data set of neuronal responses the execution time was reduced of about 45%. For balanced data sets, the platform may achieve a reduction in execution time equal to the inverse of the number of back-end computational elements.
Top1. Introduction
Parallel computing platforms provide researchers the means to apply computationally demanding algorithms to large sets of data. Although these algorithms can be applied in a sequential computing system, the use of parallelization techniques allows the substantial reduction of the execution time. One such case is the classification of neuronal responses for modeling neuronal systems and, at the bottom line, enabling the development of neural prostheses. With this goal, information being transmitted between neurons is recorded when the neuronal system is stimulated by a given stimulus. This information is communicated by biphasic electric pulses lasting a couple of milliseconds, usually known as action potentials or spikes1.
In order to understand the information being transmitted or processing mechanisms of the neuronal system, large arrays of microelectrodes are employed to perform simultaneous extracellular recordings of neuronal data, i.e. of the action potentials being communicated. The problem in performing extracellular recordings is that a single electrode typically records the action potentials originated from multiple neurons (Brown, Kass, & Mitra, 2004). This makes the analysis of the neuronal data difficult, since one does not know how many cells elicited action potentials, or even which neuron is responsible for the generation of a given action potential. To overcome this difficulty, several classification algorithms have been developed to estimate the number of active neurons and to identify which neuron generated each action potential (e.g. Tomás & Sousa, 2007; Takahashi, Anzai, & Sakurai, 2003a; Shoham, Fellows & Normann, 2003; Quiroga, Nadasdy, & Ben-Shaul, 2004).
Spike classification algorithms are computationally intensive, where the classification of spikes from a single electrode can take several hours. Moreover, in typical neural recordings, several electrodes are used to capture the action potentials from multiple locations. Nowadays arrays of 100 microelectrodes are common and it is possible to record the neuronal responses using more than one microelectrode array. This makes the spike classification a time consuming task (Quiroga, Nadasdy, & Ben-Shaul, 2004). To overcome this difficulty, one can use parallel processing in order to decrease computational time and increase the availability of the data for further analysis, which is usually the intent of researchers.
Herein is proposed a platform for mass-classification of neuronal responses, using data parallelism for reducing computational time. The proposed platform is divided into three main components: i) a front-end part for interfacing with the user, allowing him to specify the classification parameters and to check the status of the submitted jobs; ii) a middleware part for supervising the classification job and ensuring that the classification algorithm is being correctly executed; and iii) a back-end part that supplies the platform computational power. The proposed platform design has a modular approach, which guarantees that further front-end parts can be easily added to the system, such as client applications to directly interface with the middleware part. Also, other computing systems can be explored, such as those using different job schedulers. Moreover, by adopting a data parallelization approach to reduce classification time, one can easily develop new classification algorithms and integrate them into the proposed platform.
This chapter is organized as follows. Section 2 presents the classification algorithm and file format for submitting jobs for the mass-classification platform. Section 3 describes the platform architecture, detailing each of the platform’s three parts. Section 4 presents the platform implementation details, including the user interface, the mechanisms for checking the job statuses and the used back-end computing system. Experimental results are also presented, illustrating the performance of the proposed mass-classification platform. Finally, in section 5 the main conclusions are drawn.