Article Figures & Data
Figures
- Figure 1.
The spike pattern recognition model uses time-varying properties of the neural responses to recognize spiking patterns and consists of a linear chain of neurons. Sensory input (here from field L) feeds into one end of the chain, propagating along the chain using synaptic delays, turning the temporal spiking pattern into a spatial activation of chain neurons. A, Bird song (spectrogram) elicits auditory responses (spike rasters), leading to excitatory (red) and inhibitory (blue) synapses from 298 sequentially connected chained neurons (small circles) that each connect to a detector (D1). The spectrogram shows the power (red, high power; blue, low power) in different frequency bands (y-axis) as a function of time (x-axis), and the rasters show the responses of one field L neuron to 100 repeated presentations (y-axis) of the above stimulus, with each tick mark representing an elicited action potential. B, When the chain recognition circuit (with four detectors D1–D4 shown) is fed input from the field L neuron in response to the concatenation of four different songs (spectrograms), each detector fires only at the end of the correct trained song (D5–D20; data not shown).
- Figure 2.
The network learns synaptic weights through an STDP-based learning rule. A, The learning rule determines how synaptic weights change based on the relative time between presynaptic (chain neuron) and postsynaptic (detector neuron) spikes (x-axis). Closely timed spikes cause excitatory synapse strengthening (green) through long-term potentiation (LTP), whereas less closely timed spikes cause synapse weakening (orange) through long term depression (LTD). The opposite rule occurs for inhibitory synapses. B, The probability of time-to-last spike (scaled to have unity height) given a Poisson input to the chain. C, The vertical offset for the learning rule, which determines the relative levels of potentiation and depression, is set such that the ratio of positive and negative areas under the product of the learning curve (A) and the timing probability (B) is equal to 0.6 to keep learning stable. D, Chain-to-detector synaptic weights (red, excitatory; blue, inhibitory) change as learning progresses for the example site from Figure 1, and the performance of the chain recognition network (E, recording sites in gray lines) plateaus after ∼30 learning iterations.
- Figure 3.
The network performs well for different durations and performs better on more reliable and sparser inputs. A, Model performance (percentage) versus stimulus duration trained (each recording in gray, mean across sites in black). B, C, There was a strong relationship between input spike train reliability and chain recognition model performance (B; r = 0.514, p = 0.005, linear fit in dashed black) and between sparseness and performance (C; r = 0.639, p < 0.001, linear fit in dashed black). D, Normalized model recognition error (y-axis; 0% is baseline performance, −100% is no songs recognized correctly) for each site (in light gray; mean in black) increased with increasing percentage of input spike train corruption, for both shuffling spike times (left) and deleting spikes (right) for different amounts of time (x-axis). The average syllable duration for a zebra finch song (94.5 ms) is shown in dashed black.
- Figure 4.
The chain recognition network can faithfully reproduce learned spike trains in reverse. A, Three example sites (at the 75th, 50th, and 25th percentiles of recognition performance of 76, 51, and 36% for sites 1, 2, and 3) with 50 original data rasters (top) and 50 network-based replay rasters (bottom). Replay spike trains come out of the network backward and are have been time reversed here to align with the original data. B–E, The sparseness, reliability, van Rossum-based discrimination performance, and firing rate for DD, DR, and RR comparisons are shown (each unit in light gray, mean plus SE on bars), suggesting that the replay spike trains are similar to the original data.
