Soft Pattern Mining in Neuroscience Christian Borgelt Abstract While the lower-level mechanisms of neural information processing (in biological neural networks) are fairly well understood, the principles of higher-level processing remain a topic of intense debate in the neuroscience community. With many theories competing to explain how stimuli are en- coded in nerve signal (spike) patterns, data analysis tools are desired by which proper tests can be carried out on recorded parallel spike trains. This paper surveys how pattern mining methods, especially soft methods that tackle the core problems of temporal imprecision and selective participation, can help to test the temporal coincidence coding hypothesis. Future challenges consist in extending these methods, in particular to the case of spatio-temporal coding. 1 Introduction Basically all information transmission and processing in humans and animals is carried out by the nervous system, which is a network of special cells called neurons or nerve cells . These cells communicate with each other by electrical and chemical signals. While the lower-level mechanisms are fairly well understood (see Section 2) and it is widely accepted in the neuroscience community that stimuli are encoded and processed by cell assemblies rather than single cells [17, 23], it is still a topic of intense ongoing debate how exactly information is encoded and processed on such a higher level: there are many competing theories, each of which has its domain of validity. Due to modern multi-electrode arrays, which allow to record the electrical signals emitted by hundreds of neurons in parallel [9], more and more data becomes available in the form of (massively) parallel spike trains that can help to tackle the challenge of understanding higher-level neural information processing. European Centre for Soft Computing, Edificio de Investigaci´ on, c/ Gonzalo Guti´ errez Quir´os s/n, 33600 Mieres, Asturias, Spain, e-mail: christian@borgelt.net 1