A Directed-Evolution Method for Sparsification and Compression of Neural Networks with Application to Object Identification and Segmentation and considerations of optimal quantization using small number of bits

This work introduces Directed-Evolution (DE) method for sparsification of neural networks, where the relevance of parameters to the network accuracy is directly assessed and the parameters that produce the least effect on accuracy when tentatively zeroed are indeed zeroed. DE method avoids a potentially combinatorial explosion of all possible candidate sets of parameters to be zeroed in large networks by mimicking evolution in the natural world. DE uses a distillation context [5]. In this context, the original network is the teacher and DE evolves the student neural network to the sparsification goal while maintaining minimal divergence between teacher and student. After the desired sparsification level is reached in each layer of the network by DE, a variety of quantization alternatives are used on the surviving parameters to find the lowest number of bits for their representation with acceptable loss of accuracy. A procedure to find optimal distribution of quantization levels in each sparsified layer is presented. Suitable final lossless encoding of the surviving quantized parameters is used for the final parameter representation. DE was used in sample of representative neural networks using MNIST, FashionMNIST and COCO data sets with progressive larger networks. An 80 classes YOLOv3 with more than 60 million parameters network trained on COCO dataset reached 90% sparsification and correctly identifies and segments all objects identified by the original network with more than 80% confidence using 4bit parameter quantization. Compression between 40x and 80x. It has not escaped the authors that techniques from different methods can be nested. Once the best parameter set for sparsification is identified in a cycle of DE, a decision on zeroing only a sub-set of those parameters can be made using a combination of criteria like parameter magnitude and Hessian approximations.

Field-Programmable Deep Neural Network (DNN) Learning and Inference accelerator: a concept

An accelerator is a specialized integrated circuit designed to perform specific computations faster than if those were performed by CPU or GPU. A Field-Programmable DNN learning and inference accelerator (FProg-DNN) using hybrid systolic and non-systolic techniques, distributed information-control and deep pipelined structure is proposed and its microarchitecture and operation presented here. Reconfigurability attends diverse DNN designs and allows for different number of workers to be assigned to different layers as a function of the relative difference in computational load among layers. The computational delay per layer is made roughly the same along pipelined accelerator structure. VGG-16 and recently proposed Inception Modules are used for showing the flexibility of the FProg-DNN reconfigurability. Special structures were also added for a combination of convolution layer, map coincidence and feedback for state of the art learning with small set of examples, which is the focus of a companion paper by the author (Franca-Neto, 2018). The accelerator described is able to reconfigure from (1) allocating all a DNN computations to a single worker in one extreme of sub-optimal performance to (2) optimally allocating workers per layer according to computational load in each DNN layer to be realized. Due the pipelined architecture, more than 50x speedup is achieved relative to GPUs or TPUs. This speed-up is consequence of hiding the delay in transporting activation outputs from one layer to the next in a DNN behind the computations in the receiving layer. This FProg-DNN concept has been simulated and validated at behavioral-functional level.