Enhancing accuracy of compressed Convolutional Neural Networks through a transfer teacher and reinforcement guided training curriculum

Model compression techniques, such as network pruning, quantization and knowledge distillation, are essential for deploying large Convolutional Neural Networks (CNNs) on resource-constrained devices. Nevertheless, these techniques frequently lead to an accuracy loss, which affects performance in applications where precision is crucial. To mitigate accuracy loss, a novel method integrating Curriculum Learning (CL) with model compression, is proposed. Curriculum learning is a training approach in machine learning that involves progressively training a model on increasingly difficult samples. Existing CL approaches primarily rely on the manual design of scoring the difficulty of samples as well as pacing the easy to difficult examples for training. This gives rise to limitations such as inflexibility, need for expert domain knowledge and a decline in performance. Thereby, we propose a novel curriculum learning approach TRACE-CNN, i.e T ransfer-teacher and R einforcement-guided A daptive C urriculum for E nhancing C onvolutional N eural N etworks, to address these limitations. Our semi-automated CL method consists of a pre-trained transfer teacher model whose performance serves as a measure of difficulty for the training examples. Furthermore, we employ a reinforcement learning technique to schedule training according to sample difficulty rather than establishing a fixed scheduler. Experiments on two benchmark datasets demonstrate that our method, when integrated into a model compression pipeline, effectively reduces the accuracy loss usually associated with such compression techniques.