To search, Click below search items. |
| All Published Papers Search Service |
|
|
|
Title |
A Dynamic Threshold Quantization Method for Ternary Neural Networks for Low-cost Mobile Robots |
|
Author |
Son Ngoc Truong |
| Citation |
Vol. 20 No. 2 pp. 16-20 |
|
Abstract |
A ternary neural network is an optimized model of deep neural networks that can be deployed on low-cost computers for mobile applications such as Internet of Things end-node devices and mobile robots. Quantization of real valued synaptic weights is performed during the training process to achieve the ternary synaptic weights of ?1, 0, and +1 for the ternary neural network. The conventional quantization method uses the sign function for classifying the synaptic weights to ?1 and 1 in the binary neural networks. For the ternary neural network, a fixed threshold is used for quantization. In this paper, we propose a dynamic threshold for quantization for ternary neural networks. The threshold value is calculated based on the distribution of real valued synaptic weights to obtain the same number of negative synaptic weights (?1), zeros, and positive synaptic weights (+1). By doing this, the accuracy and training time is improved significantly. For MNIST handwritten recognition, training the ternary neural network using the proposed dynamic threshold quantization method is 2.5× faster than using the convention method. The recognition rate of a ternary neural network with the proposed quantization method shows better recognition rate by 2% compared to the ternary neural network with fixed threshold quantization method. For the application of speech recognition, the ternary neural network is trained and evaluated using Google Speech Command dataset. The ternary neural network with the proposed quantization method achieves the recognition rate of 8% higher than the fixes threshold quantization method. The training time is also reduced by half when using the proposed quantization method. The ternary neural network with the proposed quantization method is useful for low-cost computers which can be deployed on the Internet of Things end-node devices and mobile robots. |
|
Keywords |
Artificial neural network, deep learning, speech recognition, image recognition, ternary neural network. |
|
URL |
