Seminar 2022-11-28

[ArieRS]

Seminar 2022-11-29

Progress

• Symposium Experiment Result
  • Illustration of research methodology
  • 
  • Dataset : WLASL (https://dxli94.github.io/WLASL/)
  • Input: Keypoint extraction of the sign language
  • Method: LSTM, LSTM + Dropout, BiLSTM, BiLSTM+Dropout, GRU, GRU+Dropout, BiGRU, BiGRU+Dropout

Method	Architecture	Accuracy	Time (Sec)
LSTM1	2 LSTM + 3 Dense layers	44%	0.38
LSTM2	2 LSTM + Dropout + 3 Dense layers	33%	0.39
BiLSTM1	2 BiLSTM + 3 Dense layers	33%	0.65
BiLSTM2	2 BiLSTM + 2 Dropout + 3 Dense layers	33%	0.65
GRU1	3 {GRU + Dense} + BatchNorm + 1 Dense layer	56%	0.44
GRU2	3 {GRU + Dropout + Dense} + BatchNorm + 1 Dense layer	56%	0.45
BiGRU1	3 {BiGRU + Dense} + BatchNorm. + 1 Dense layer	56%	0.75
BiGRU2	3 {BiGRU + Dropout + Dense} + BatchNorm. + 1 Dense layer	50%	0.69

• Reservoir Computing Progress
  • Successfully feed the extracted key point to Multivariate Reservoir Computing

Method	Architecture	Accuracy	Time (Sec)
Multivariate RC	450 Reservoir	56%	0.19
Multivariate RC	550 Reservoir	67%	0.26

Next Plan

• Creating research posters for symposiums - Deadline 8th December 2022
• Continuing to explore multivariate reservoir computing
• Implementing VAE or another dimensional reduction method
• Drafting journal or conference paper for reservoir computing
• Trying out existing implementations to gain insight into gesture recognition
• Try to implement Reservoir Computing
• Drafting symposium paper
• Adding literature study