Algorithm proves useful for analysis of neuronal data
Generalization performance measured as the number of misclassified noisy patterns over the total number of tested noisy patterns. A complete explanation can be found in the original article.
An article by researchers from the Bernstein Focus Neurotechnology / Bernstein Center Freiburg, appearing in the journal Frontiers in Computational Neuroscience, presents a new approach to the problem of classifying temporal patterns in the spiking activity of nerve cells. The scientists devised a novel supervised learning rule that is based on the widely used concept known as “Support Vector Machines”. They demonstrated how this technique can be used to determine the synaptic weights of a standard model neuron, which after training very effectively classifies spike patterns – and proves to possess other uses in computer technology.
Spike pattern classification has become an important topic in several fields of research. For example, electrophysiological recordings from neural networks on Multi-Electrode Arrays (MEA) provide spike patterns distributed in space and time that remain difficult to understand. Furthermore, theoretical studies have shown that spiking nerve cells are better suited for fast information processing than neurons using graded signals. Finally, electronic devices based on impulses and temporal coding strategies have useful properties that can be exploited for sensors and neuronal prostheses. Recent research on novel computer architectures also explores spike based processors that consume less power than conventional chips.
An important aspect of neural processing is the temporal integration of postsynaptic excitatory and inhibitory signals. The result of this integration determines the neural response in the form of brief electrical impulses, the so-called action potentials or spikes. In this scenario, the transfer between input signals and output signals in a nerve cell is determined by the amplitude and the time course of synaptic events, dubbed “synaptic weights”. In their research article, Maxime Ambard and Stefan Rotter offer a new way to determine the synaptic weights of a leaky integrate-and-fire neuron model in a way that is optimal for spike pattern classification. They used a supervised learning rule to accomplish this goal. The rule is based on so-called “Support Vector Machines”, a specific mathematical procedure that is used in the field of machine learning to solve difficult classification tasks. The scientists from Freiburg compared their model’s performance to classify spike patterns with several other methods trying to solve the same problem.
As their analysis showed, the new algorithm also performed quite well in generalization tasks. They could further determine that the ability to separate spike patterns depended on a relation between the exact shape of postsynaptic potential and the duration of spike patterns. From their insights into the performance of their method, Ambard and Rotter propose a particular shape that is particularly well-suited for fast computations, and therefore can be used in time critical implementations of the algorithm based on specialized software or hardware.
Original publication:
Ambard M and Rotter S (2012) Support vector machines for spike pattern classification with a leaky integrate-and-fire neuron. Front. Comput. Neurosci. 6:78. doi: 10.3389/fncom.2012.00078