Figure 6
(a) A schematic of the decoding scheme for sequences. (b) Decoding accuracy as a function of sequence position \(\ell \) along sequences of length L retrieved from memory shown for several values of L. \(N=2500\), \(\alpha =0.05\), and the average is computed over 100 trials. (c) A schematic of the energy landscape for two temporal sequences containing a single common element which are encoded as sequences of attractors using the scheme in34 (discussed in Section 4 of the Supplementary Material). (d) The average overlap of the network state \(\langle m \rangle \) with each attractor in the sequence is shown as a function of retrieval time for a network of size \(N=1000\) storing two sequences containing a single element in common. The parameters \(\tau =8\) and \(\lambda =2.5\) are used and the average is computed over 1000 trials. (e) The average overlap of the estimator \(\hat{a}_{\ell }\) (normalized by \(\langle \hat{a}_{2}\cdot a_{2}\rangle \)) with the correct Dictionary item at each position in the sequence for a network of size \(N=1000\) storing two sequences containing a single element in common. The average is computed over 10, 000 trials. (f) Decoding accuracy as a function of sequence position \(\ell \) for \(P=\alpha N\) sequences for which each neighboring sequence contain \(L_{s}\) common elements with the previous one. \(N=2500\), \(\alpha =0.05\) and the average is computed over 100 trials.
