Jump-GRS: a multi-phase approach to structured pruning of neural networks for neural decoding.

Neural decoding, an important area of neural engineering, helps to link neural activity to behavior. Deep neural networks (DNNs), which are becoming increasingly popular in many application fields of machine learning, show promising performance in neural decoding compared to traditional neural decoding methods. Various neural decoding applications, such as brain computer interface applications, require both high decoding accuracy and real-time decoding speed. Pruning methods are used to produce compact DNN models for faster computational speed. Greedy inter-layer order with Random Selection (GRS) is a recently-designed structured pruning method that derives compact DNN models for calcium-imaging-based neural decoding. Although GRS has advantages in terms of detailed structure analysis and consideration of both learned information and model structure during the pruning process, the method is very computationally intensive, and is not feasible when large-scale DNN models need to be pruned within typical constraints on time and computational resources. Large-scale DNN models arise in neural decoding when large numbers of neurons are involved. In this paper, we build on GRS to develop a new structured pruning algorithm called jump GRS (JGRS) that is designed to efficiently compress large-scale DNN models.On top of GRS, JGRS implements a 'jump mechanism', which bypasses retraining intermediate models when model accuracy is relatively less sensitive to pruning operations. Design of the jump mechanism is motivated by identifying different phases of the structured pruning process, where retraining can be done infrequently in earlier phases without sacrificing accuracy. The jump mechanism helps to significantly speed up execution of the pruning process and greatly enhance its scalability. We compare the pruning performance and speed of JGRS and GRS with extensive experiments in the context of neural decoding.Our results demonstrate that JGRS provides significantly faster pruning speed compared to GRS, and at the same time, JGRS provides pruned models that are similarly compact as those generated by GRS.In our experiments, we demonstrate that JGRS achieves on average 9%-20% more compressed models compared to GRS with 2-8 times faster speed (less time required for pruning) across four different initial models on a relevant dataset for neural data analysis.