Purchase this article with an account.
Ulrich Weidenbacher, Heiko Neumann; The first spike counts: A model for STDP learning pose specific representations for estimating view direction. Journal of Vision 2008;8(6):161. doi: https://doi.org/10.1167/8.6.161.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Perceptual investigations demonstrate that a person's view direction can be reliably estimated on the basis of perceived head orientation and eye gaze (Langton et al., TICS 4, 2000). VanRullen & Thorpe (Vision Res.42, 2002) proposed a feedforward model which employs neural rank order coding to achieve sparse representations of objects (faces). The unsupervised development of a neural face representation can be modeled by utilizing STDP learning (Masquelier & Thorpe, PLoS Comp. Biol. 3, 2007). STDP learning is a highly evidenced mechanism of Hebbian learning based on the temporal order of spike generation of the pre- and postsynaptic neuron at a particular synapse (Bi & Poo, J.Neurosci.18, 1998). We applied a model of STDP to automatically learn different face poses. As input we sequentially probed the model by images of faces in different pose conditions. Neural activities were generated by convolving the image by differently scaled orientation selective Gabor filters which responses subsequently undergo shunting competition. Activations (resembling the spike rate of neurons) were converted into spike firing latencies (temporal spike order) followed by lateral inhibition. Further processing is limited to a maximum number of neurons that fire early. The spike times were then transferred to a category layer where prototype neurons are dynamically allocated and their synaptic weights are trained by applying the STDP learning rule.We demonstrate that the model automatically finds an appropriate number of output neurons depending on the statistical regularities of the input patterns. Furthermore, we show that these neurons become selective to different face poses that were repeatedly given as input to the model. In conclusion, rank order coding in combination with STDP learning is a very efficient and rapid scheme that can be used to robustly learn intermediate level representations of face patterns, such as face poses. The resulting sparse representation of prototypes enables reliable head pose estimates in mutual visual communication.
This PDF is available to Subscribers Only