The parallel learning in neural networks can greatly shorten the training time. Its prior efforts were mostly limited to distributing inputs to multiple computing engines. It is because the gradient descent algorithm in the neural network training is inherently sequential. This paper proposes a novel CNN parallel training method for image recognition. It overcomes the sequential property of the gradient descent and enables the parallel training with the speculative backpropagation. We found that the Softmax and ReLU outcomes in the forward propagation for the same labels are likely to be very similar. This characteristic makes it possible to perform the forward and backward propagation simultaneously. We implemented the proposed parallel model with CNNs in both software and hardware, and evaluated its performance. The parallel training reduces the training time by 34% in CIFAR-100 without the loss of the prediction accuracy compared to the sequential training. In many cases, it even improves the accuracy.

Continual learning is gaining traction these days with the explosive emergence of deep learning applications. Continual learning suffers from a severe problem called catastrophic forgetting. It means that the trained model loses the previously learned information when training with new data. This paper proposes two novel ideas for mitigating catastrophic forgetting: Speculative Backpropagation (SB) and Activation History (AH). The SB enables performing backpropagation based on past knowledge. The AH enables isolating important weights for the previous task. We evaluated the performance of our scheme in terms of accuracy and training time. The experiment results show a 4.4% improvement in knowledge preservation and a 31% reduction in training time, compared to the state-of-the-arts (EWC and SI).