An Optimized Cosine Similarity-Based Attention Gate for Temporal Sequence Patterns Recognition

Abstract The attention gate is often employed in various applications, such as recommender systems, to determine the correct context of the input data. Adopting the attention gate has been proven to improve prediction accuracy successfully. However, for a temporal sequence problem such as sensor-based Sign Language Recognition (SLR), it is still challenging to integrate the attention gate into the solution since the processing input is only in the form of a sine waveform. In our work, we propose an optimized cosine similarity-based attention gate to manipulate the sine wave signal while improving the recognition of temporal sequence data. We also evaluate and compare the various distance calculations and determine the best one for the SLR application. Adopting a distance-based attention gate has successfully achieved the recognition accuracy of 99% while reducing the error rate to 5%.