STDP Installs in Winner-Take-All Circuits an Online Approximation to Hidden Markov Model Learning
Figure 4
Mixed selectivity in networks of multiple interconnected WTA circuits.
(A,B) Mean firing rate of the circuit neurons for evoked activity during pattern a in sequence AB-delay-ab (A) and BA-delay-ba (B). A threshold of 10 Hz (dashed line) was used to distinguish between neurons that were active or inactive during the pattern. Firing rates of neurons that were not context selective are shown in green, that of neurons selective for starting sequences AB and BA are shown in red and blue, respectively. Neurons that did not fall in one of these groups are not shown. Spike trains of one context selective (C) and one non-selective (D) neuron are presented for spontaneous completion of sequence AB-delay-ab (upper) and BA-delay-ba (lower) (cue phase is not shown). Spike raster plots over 20 trial runs and corresponding averaged neural activity (PETH) are shown. The two neurons encode the input on different levels of abstraction. The neuron in panel (D) shows context cell behavior, since it encodes pattern a only if it occurs in the context of sequence ab. During ba it remains (almost) perfectly silent. The neuron in (C) is not context selective, but nevertheless fires reliably during the time slot of pattern a during the free run by integrating information from other (context selective) neurons. It belongs to a WTA circuit with 15 neurons, for which the network state projection is shown in panel (E). (E,F) Linear projection of the network activity during the delay phase to the first two components of the jPCA, for a single WTA circuit with 15 neurons (E) and for the whole network (F). 10 trajectories are plotted for each sequence (AB-delay-ab red, BA-delay-ba green, CD-delay-cd blue, DC-delay-dc yellow). The dots at the beginning of each line, indicate the onsets of the delay state, i.e. the beginning of the trajectories. The plots have arbitrary scale. The projection of the WTA circuit in (E) does not allow a linear separation between all four sequences, whereas the activity of the whole network (F) clusters into four sequence-specific regions. The network neurons use this state representation to modulate their behavior during spontaneous activity.