Skip to Main content Skip to Navigation

Approximate computing for embedded machine learning

Abstract : Convolutional Neural Networks (CNNs) have been extensively used in many fields such as image recognition, video processing, and naturallanguage processing. However, CNNs are still computational-intensive and resource-consuming. They are often constrained by the limit performanceand memory when deployed on embedded systems. This PhD research project aims at proposing CNNs which are more suitable for embedded systems withlow computing resources and memory requirements. Based on literature review, we propose three methods to accelerate the operation of neural networks : Selective Binarization, Quad-Approx Network and Min- ConvNets. Selective Binarization combines layers with different precisions in CNNs to achieve an acceptable speed and accuracy. As well an FPGA based hardware accelerator is proposed for these optimized structures. With the proposed signed PArameterized Clipping acTivation Function (signed PACT), the CNNs are quantized into 3 bits, and then a loss-less network is established by using approximate multiplier, which is named Quad-Approx Network. In addition to acceleration, what is more valuable is that Quad-Approx shows that CNNs are certain fault tolerance systems, which leads us to propose the MinConvNets. MinConvNet is a set of multiplication-less CNNs whose multiplications are replaced by approximate operations. MinConvNet can achieve negligible loss of prediction compared to exact image classification networks through transfer learning, meanwhile the multiplication which is more resource consuming to implement is replaced by easier implemented operations. Human is ushering the era of the artificial intelligence. In the meantime, the Internet of Things (IoT) makes our lives more convenient. These works bring more complex intelligent algorithms into the edge devices and helps us to create the era of Artificial intelligent Internet of Things (AIoT).
Complete list of metadata
Contributor : Abes Star :  Contact
Submitted on : Monday, June 28, 2021 - 2:54:09 PM
Last modification on : Tuesday, June 29, 2021 - 3:34:42 AM
Long-term archiving on: : Wednesday, September 29, 2021 - 7:38:12 PM


Version validated by the jury (STAR)


  • HAL Id : tel-03272594, version 1



Xuecan Yang. Approximate computing for embedded machine learning. Electronics. Institut Polytechnique de Paris, 2021. English. ⟨NNT : 2021IPPAT005⟩. ⟨tel-03272594⟩



Record views


Files downloads