A low‐cost compensated approximate multiplier for Bfloat16 data processing on convolutional neural network inference

Abstract

This paper presents a low‐cost two‐stage approximate multiplier for bfloat16 (brain floating‐point) data processing. For cost‐efficient approximate multiplication, the first stage implements Mitchell's algorithm that performs the approximate multiplication using only two adders. The second stage adopts the exact multiplication to compensate for the error from the first stage by multiplying error terms and adding its truncated result to the final output. In our design, the low‐cost multiplications in both stages can reduce hardware costs significantly and provide low relative errors by compensating for the error from the first stage. We apply our approximate multiplier to the convolutional neural network (CNN) inferences, which shows small accuracy drops with well‐known pre‐trained models for the ImageNet database. Therefore, our design allows low‐cost CNN inference systems with high test accuracy.

MLCNN‐COV: A multilabel convolutional neural network‐based framework to identify negative COVID medicine responses from the chemical three‐dimensional conformer

“…Lightweight depthwise convolutions are used in the intermediate expansion layer as a source of nonlinearity to filter features. The MobileNetV2 architecture includes a 32‐filter with an initial fully convolution layer as well as 19 residual bottleneck layers [40, 41]. DenseNet201: DenseNet is based on the notion that convolutional networks may be trained to be significantly deeper, more precise, and more effective if they have shorter connections between layers that are close to the input and those that are close to the output.…”

“…The most used deep learning model for image classification is CNN and transfer-learning models [38,39]. The two parts of the presented MLCNN-COV framework are presented in Figure 5.…”

Energy-efficient neural network learning with accuracy-adjustable floating-point multiplier

“…In precision scaling [1], we use fewer bits to represent numeric values rather than executing all the required mathematical operations with the full representation. Several standards for the floating-point presentation recently appeared: IEEE 754-2019 for half-precision [2], posit format with dynamic range and mantissa [3], and Google's bfloat16, targeting the machine-learning workloads [4]. Storing the numeric values with fewer bits reduces the size of arithmetic circuits and their complexity.…”