, 2004). For a variety of animal species and for different modalities it has been demonstrated BTK inhibitor solubility dmso that single neurons respond in a temporally sparse manner (Reinagel, 2001, Jadhav et al., 2009, Olshausen et al., 2004 and Hromádka et al., 2008) when stimulated

with natural time-varying input. In the mammal this is intensely studied in the visual (Dan et al., 1996, Vinje and Gallant, 2000, Reinagel and Reid, 2002, Yen et al., 2007, Maldonado et al., 2008, Haider et al., 2010 and Martin and Schröder, 2013) and the auditory (Hromádka et al., 2008, Chen et al., 2012 and Carlson et al., 2012) pathway as well as in the rodent whisker system (Jadhav et al., 2009 and Wolfe et al., 2010). Sparseness increases across sensory

processing levels and is particularly high in the neocortex. Individual neurons emit only a few spikes positioned at specific instances during the presentation of a time-varying input. Repeated identical stimulations yield a high reliability and temporal precision of responses (Herikstad et al., 2011 and Haider et al., 2010). Thus, single Rucaparib price neurons focus only on a highly specific spatio-temporal feature from a complex input scenario. Theoretical studies addressing the efficient coding of natural images in the mammalian visual system have been very successful. In a ground breaking study, Olshausen et al. (1996) learned a dictionary of features for reconstructing a large set of natural still images under Reverse transcriptase the constraint of a sparse code to obtain receptive fields (RFs), which closely resembled the physiologically measured RFs of simple cells in the mammalian visual

cortex. This approach was later extended to the temporal domain by van Hateren and Ruderman (1998), learning rich spatio-temporal receptive fields directly from movie patches. In recent years, it has been shown that a number of unsupervised learning algorithms, including the denoising Autoencoder (dAE) (Vincent et al., 2010) and the Restricted Boltzmann Machine (RBM) (Hinton and Salakhutdinov, 2006, Hinton et al., 2012 and Mohamed et al., 2011), are able to learn structure from natural stimuli and that the types of structure learnt can again be related to cortical RFs as measured in the mammalian brain (Saxe et al., 2011, Lee et al., 2008 and Lee et al., 2009). Considering that sensory experience is per se dynamic and under the constraint of a temporally sparse stimulus representation at the level of single neurons, how could the static RF model, i.e. the learned spatial feature, extend into the time domain? Here we address this question with an unsupervised learning approach using RBMs as a model class.