US 12,353,990 B2
Time-window based attention long short-term memory network of deep learning
Toshiya Iwamori, Tokyo (JP); Akira Koseki, Yokohama (JP); Hiroki Yanagisawa, Tokyo (JP); and Takayuki Katsuki, Tokyo (JP)
Assigned to International Business Machines Corporation, Armonk, NY (US)
Filed by International Business Machines Corporation, Armonk, NY (US)
Filed on Jul. 12, 2020, as Appl. No. 16/926,741.
Prior Publication US 2022/0013239 A1, Jan. 13, 2022
Int. Cl. G06N 3/08 (2023.01); G16H 50/30 (2018.01); G16H 50/50 (2018.01); G16H 50/70 (2018.01); G16H 70/60 (2018.01)
CPC G06N 3/08 (2013.01) [G16H 50/30 (2018.01); G16H 50/50 (2018.01); G16H 50/70 (2018.01); G16H 70/60 (2018.01)] 18 Claims
OG exemplary drawing
 
1. A computer-implemented method for a time-window based attention long short-term memory network (TW-LSTM network), the method comprising:
a time-window based attention long short-term memory network (TW-LSTM network) splitting elapsed time into a predetermined number of time windows, wherein the elapsed time spans from a current cell state back to a predetermined number of previous cell states, wherein respective ones of the windows have respective numbers of cell states;
the TW-LSTM network calculating average values of the previous cell states in the respective ones of the time windows;
the TW-LSTM network setting the average values as aggregated cell states for the respective ones of the time windows;
the TW-LSTM network generating attention weights for the respective ones of the time windows;
the TW-LSTM network calculating a new previous cell state, based on the aggregated cell states and the attention weights, wherein the new previous cell state is a weighted summation of the aggregated cell states;
the TW-LSTM network updating the current cell state, based on the new previous cell state; and
the TW-LSTM network predicting a target variable for sequential data with time irregularity, based on a current hidden state that is generated from an updated current cell state.