Feed-forward processing of speech data.

<p>(<b>A</b>) Histogram of subthreshold potentials for the cell illustrated in <a href="http://www.ploscompbiol.org/article/info:doi/10.1371/journal.pcbi.1003272#pcbi-1003272-g003" target="_blank">Figure 3</a>, accumulated over all test-set data after learning with three different plasticity rules. (<b>B</b>) Distribution of skewness for 4500 neurons trained similarly from random initial weights. Skewness after Convallis training is significantly higher than after rate constraint or rcSTDP, but rcSTDP and rate constraint do not differ. (<b>C</b>) Mean rate response of the example neuron to all digits. Errors bar show s.e.m. (<b>D</b>) The strength of tuning for each neuron was summarized by an F-statistic that measured the selectivity of its spike counts for particular digits (see <a href="http://www.ploscompbiol.org/article/info:doi/10.1371/journal.pcbi.1003272#s4" target="_blank">Materials and Methods</a>). The main graph shows an histogram of tuning strength across the simulated population for the 3 learning rules and the raw cochleogram input, while the inset shows mean and standard error. Note that while the Convallis rule produces sharper tuning than the cochleogram inputs, rate constraint alone and rcSTDP produce weaker tuning. (<b>E</b>) To evaluate the ability of these rules to perform unsupervised learning, the spike count responses of up to 4500 cells were used as input to a linear classifier trained to distinguish digits. (<b>F</b>) Mean classification performance as a function of the number of unsupervised neurons. (Errors bars show s.e.m over 10 independent runs of the analysis). The left axis marks the number of neurons in the cochleogram representation, and the horizontal dashed line indicates classification from the raw cochleogram. Note that Convallis outperforms the raw cochleogram, populations trained by rate constraint, and populations trained by rcSTDP for all numbers of neurons.