A DNN Optimization Framework with Unlabeled Data for Efficient and Accurate Reconfigurable Hardware Inference

Abstract

Open-source deep-learning frameworks are prevalent in designing, training, and deploying deep neural networks (DNNs) on general-purpose computing devices, such as CPU, GPU, and DSP. However, for custom-designed reconfigurable hardware accelerators, there is no existing universal framework, capable of optimizing DNN deployment configuration and guiding the hardware design with specific accuracy and efficiency requirements. In the paper, we proposed a crossplatform framework, which can convert deep-learning models from popular open-source frameworks to intermediate representation and optimize weight/activation dynamic ranges and quantization strategy to achieve better efficiency and accuracy based on a baseline reference design of hardware accelerator. With a few unlabeled data, the proposed framework can analyze the statistical inference information, compare different bit-width impacts, and optimize network structure. We further illustrate the detailed experiment results using the framework, showing mAP and top-1 accuracy loss is less than 1.5% and 1.2% with 12-bit and 8-bit activation-constrained quantization schemes respectively for object detection and image classification.

A Compilation Framework for SRAM Computing-in-Memory Systems with Optimized Weight Mapping and Error Correction

“…In addition, limited research has focused on profiling both the bit-flexible and signed-unsigned reconfigurable mapping, which could potentially benefit the broad application of SRAM-CIM systems. Bit flexibility is important for SRAM CIM to balance the algorithm accuracy and hardware throughput [25]. Typical algorithms of CNN, DSP, and DIP yield different requirements of signed/unsigned data format.…”

“…Thus low power consumption [36] can be achieved with acceptable accuracy loss. This quantizer is based on the quantization algorithm [25] proposed by Chen et al, which is a post-training quantization strategy with min-max symmetric scaling parameters. Additional unsigned quantization is supported to improve algorithm accuracy if inputs or weights are positive in DSP, DIP and CNN.…”

Optimal planning and allocation of Plug-in Hybrid Electric Vehicles charging stations using a novel hybrid optimization technique

“…The pseudo code of the developed AT-AO framework is presented below in Algorithm 1 and its flowchart is depicted in Fig 2. (20) and Eq (21); Employing Eq (24), compute its K neighbors. Utilizing Eq (16) and Eq (17), estimate the interaction force F p and the constraint force correspondingly.…”

“…The simplest way to overcome this problem is to use charging synchronization. The idea is that EVs should submit data, such as battery SoC so that a system may prioritize charging demands and decide which EV should charge during this time slot, while delaying other demands to future time slots [18][19][20][21].…”