Learning precise spatiotemporal sequences via biophysically realistic learning rules in a modular, spiking network
- Neurobiology and Anatomy, University of Texas Medical School at Houston, United States
- Applied Physics, Rice University, United States
Figures
Figure 1
Sequence representation in networks.
(a) A network composed of different populations of cells, each population is activated by a specific stimulus, and there are plastic connections between and within these populations. Initially these …
Figure 2
Microcircuit learns time intervals.
(a) Mean firing rates of Timer (black), Messenger (red), and inhibitory populations (light blue) in a microcircuit before learning (top) and after learning (bottom) to represent an 1100 ms interval. …
Figure 3 with 6 supplements
Sequence learning and recall.
(a) Network of 12 columns, each containing a core neural architecture (CNA) microcircuit selective for a different stimulus. Columns containing microcircuits responding to blue, red, green, and …
Figure 3—figure supplement 1
Input layer dynamics.
(a) Simulations showing response of input layer units to 400 ms stimulus (fixed spot size, seven degrees). The input is approximated as a 50 ms pulse of Poisson spikes. This is the approximation …
Figure 3—figure supplement 2
Accuracy of learning and recall.
A network is trained to a sequence of four elements, each 700 ms in duration. Owing to stochastic nature of spiking network, reported times can fluctuate from presentation to presentation, and from …
Figure 3—figure supplement 3
Learning of different sequences.
Left, identity and order of stimuli shown during training. Right, mean firing rate of network after training, upon stimulation of first column in sequence. (a) Blue, green, red, and orange columns …
Figure 3—figure supplement 4
Spiking statistics in learned network.
(a) Spike raster of network response to stimulation of first column (light blue bar), after learning a sequence of stimuli (500, 750, 500, and 1250 ms for columns 1, 2, 3, and 4, respectively). …
Figure 3—figure supplement 5
Eight element sequence recall.
Recall after learning a sequence of eight elements, each with duration 700 ms. Only the first element is stimulated. Notice that because of stochasticity, some elements (1 and 8) underreport their …
Figure 3—figure supplement 6
Rate-based learning and recall.
Recreation of Figure 3 from the main text, but using the rate-based formulation described in 'Materials and methods'. Each population of previously spiking neurons (e.g. red Timers) is now …
Figure 4 with 2 supplements
Change in connectivity patterns resulting from learning.
(a) Before, (b) during, and (c) after learning a sequence. Left, view of columnar structure and learned intercolumnar connections. Dotted box indicates region shown in side view, middle. Middle, the …
Figure 4—figure supplement 1
Recurrent learning evolution.
Top row: firing rates of Timer (light colors) and Messenger (dark colors) populations for the first two columns over the course of learning. Bottom row: eligibility traces corresponding to the …
Figure 4—figure supplement 2
Feed-forward learning evolution.
Top row: firing rates of Timer (light colors) and Messenger (dark colors) populations for the first two columns. Bottom row: eligibility traces corresponding to the feed-forward weights between the …
Figure 5 with 2 supplements
Robustness to core neural architecture (CNA) weight changes.
Firing rates of four columns, after learning a four element sequence, each of 700 ms duration. Only the first element is stimulated for recall. Before learning, static CNA weights WEEMT and WEIMT …
Figure 5—figure supplement 1
Robustness of parameter randomization.
Left, a two-column network learns a two-element sequence (500 ms, 500 ms) over 100 different learning epochs. The mean recalled time (bar) and standard deviation of the recalled times (whiskers) are …
Figure 5—figure supplement 2
Sequence learning with 20 ms excitatory time constant.
(a) A network with 20 ms excitatory time constants recalls (only first element stimulated) a learned four element sequence of 500 ms each. Sequence learning is successful and network timing is …
Figure 6
Non-Markovian sequence learning and recall.
Three-stage network. Two sequentially activated columns (2–3) learn to connect to each other through a reservoir and sparse pattern net. At time t, Messenger cells from column 2 are active and act …
Figure 7
Non-Markovian sequence learning and recall.
Mean firing rates for Timer cells (light colors) and Messenger cells (dark colors) of four different columns during different stages of learning (before, first trial of learning, last trial of …
Figure 8 with 1 supplement
Recall of two overlapping sequences.
Mean firing rates for Timer cells (light colors) and Messenger cells (dark colors) during recall of two sequences. Both blue-red-orange (BRO) and green-red-purple (GRP) have been stored in the …
Figure 8—figure supplement 1
Robustness in non-Markovian recall.
A three-stage network trained on two non-Markovian sequences (BRO and GRP) recalls the two sequences with and without a perturbation to the initial state of the reservoir. (a) The trained network is …
Figure 9
Explicit microcircuit structure.
(a,b) Two examples of complete microcircuit structure displayed in laminar architecture of cortical columns. Dashed lines represented learned connections, while continuous lines represent fixed …
Additional files
-
Supplementary file 1
Table of main model parameters.
For full code, see http://modeldb.yale.
- https://cdn.elifesciences.org/articles/63751/elife-63751-supp1-v2.docx
-
Supplementary file 2
Table of reservoir, sparse net, and rate-based model parameters.
- https://cdn.elifesciences.org/articles/63751/elife-63751-supp2-v2.docx
-
Transparent reporting form
- https://cdn.elifesciences.org/articles/63751/elife-63751-transrepform-v2.pdf
