A Recurrent Self-Organizing Map for Temporal Sequence Processing

Tracking non-stationary dynamical system phase using multi-map and temporal self-organizing architecture

“…Some other works [3] focused on setting up an on-map representations to inputs, and used the difference in representations on the map to compare input temporal sequences. Reservoir computing approach [4] aims basically to set up a mapping between the input phase space and the reservoir state space.…”

Extracting Finite Structure from Infinite Language

“…In contrast with the majority of supervised connectionist models in the literature, STORM is based on an unsupervised recurrent SOM [15] and operates using a discrete state-space.…”

“…STORM's predictions are made by utilizing the locational representational values used in its context vector. As further explained in [15], the winning neuron for an input is the neuron whose weight vector best matches both the input symbol and the context representation of the last winning neuron's location. STORM predicts the next symbol by finding the neuron whose context representation best matches that of the current winning neuron (i.e.…”

“…STORM is a recurrent SOM [15] that acts as a temporal associative memory, initially producing a localist-based memorization of input sequences. The model's rule-based mechanism then exploits similarities between the future context of memorized sequences and the future context of input sequences.…”

“…By separately representing the column and row of the previous winning neuron in the context vector, the recurrency mechanism creates a 2D representation of the neuron's location. Further details of the recurrency mechanism, along with its advantages, are provided in [15].…”