Search

Processing-in-memory (PIM), an increasingly studied neuromorphic hardware, promises orders of energy and throughput improvements for deep learning inference. Leveraging the massively parallel and efficient analog computing inside memories, PIM circumvents the bottlenecks of data movements in conventional digital hardware. However, an extra quantization step (i.e. PIM quantization), typically with limited resolution due to hardware constraints, is required to convert the analog computing results into digital domain. Meanwhile, non-ideal effects extensively exist in PIM quantization because of the imperfect analog-to-digital interface, which further compromises the inference accuracy. In this paper, we propose a method for training quantized networks to incorporate PIM quantization, which is ubiquitous to all PIM systems. Specifically, we propose a PIM quantization aware training (PIM-QAT) algorithm, and introduce rescaling techniques during backward and forward propagation by analyzing the training dynamics to facilitate training convergence. We also propose two techniques, namely batch normalization (BN) calibration and adjusted precision training, to suppress the adverse effects of non-ideal linearity and stochastic thermal noise involved in real PIM chips. Our method is validated on three mainstream PIM decomposition schemes, and physically on a prototype chip. Comparing with directly deploying conventionally trained quantized model on PIM systems, which does not take into account this extra quantization step and thus fails, our method provides significant improvement. It also achieves comparable inference accuracy on PIM systems as that of conventionally quantized models on digital hardware, across CIFAR10 and CIFAR100 datasets using various network depths for the most popular network topology., Comment: 25 pages, 12 figures, 8 tables

This report covers an intelligent decision support system (IDSS), which handles an efficient and effective way to rapidly inspect containerized cargos for defection. Defection is either cargo exposure to radiation, physical damages such as holes, punctured surfaces, iron surface oxidation, etc. The system uses a sorting array triangulation technique (SAT) and surface damage detection (SDD) to conduct the inspection. This new technique saves time and money on finding damaged goods during transportation such that, instead of running $n$ inspections on $n$ containers, only 3 inspections per triangulation or a ratio of $3:n$ is required, assuming $n > 3$ containers. The damaged stack in the array is virtually detected contiguous to an actually-damaged cargo by calculating nearby distances of such cargos, delivering reliable estimates for the whole local stack population. The estimated values on damaged, somewhat damaged and undamaged cargo stacks, are listed and profiled after being sorted by the program, thereby submitted to the manager for a final decision. The report describes the problem domain and the implementation of the simulator prototype, showing how the system operates via software, hardware with/without human agents, conducting real-time inspections and management per se., Comment: This article is a report on a developed IDSS system/prototype aimed to be published for a journal conference proceedings and/or a full paper under Computer Science and Software Engineering categories. 28 pages; 10 Figures including graphs; 5 tables; presentation file is available at http://web.uvic.ca/~phibal12/Presentations/IDSS_proj.pptx Ask authors for full code and/or other files

This article is about the architecture of a lossless wavelet filter bank with reprogrammable logic. It is based on second generation of wavelets with a reduced of number of operations. A new basic structure for parallel architecture and modules to forward and backward integer discrete wavelet transform is proposed., Comment: 7 pages, The ACM proceedings of CompSysTech 2003