Comparative Study of Optimization Algorithm in Deep CNN-Based Model for Sign Language Recognition

The fundamental part of the neural network is the learning rate, and the strategy of adopting the learning process in a neural network is carried out using optimization algorithms or optimizers. This optimization algorithm helps us produce better results to the model by changing the parameters like bias and weights, i.e., it helps us maximize or minimize the error function and depends on the learnable parameters. In this paper, we examine how an End-to-End CNN model named ASLNET recognizes the alphabets of the American sign language using various optimizers such as Stochastic Gradient Descent (SGD), Root-Mean-Square propagation (RM-Sprop), Adaptive Gradient Algorithm (Adagrad), Adaptive Delta (Adadelta),Adaptive Moment Estimation (Adam), Adam with Nesterov Momentum (Nadam), LookAhead and Rectified Adam (RAdam). To avoid the overfitting issues, traditional data augmentation techniques are used to compare our model with data augmentation and without augmentation with these optimizers. Among these, LookAhead and RAdam are the most recently developed. The experiment is conducted on 2 NVIDIA TESLA P100 GPUs of batch size 64, and the investigation was based on benchmark ASL Finger Spelling dataset.