Modular Neural Networks for Low-Power Image Classification on Embedded Devices

Abstract

Embedded devices are generally small, battery-powered computers with limited hardware resources. It is difficult to run deep neural networks (DNNs) on these devices, because DNNs perform millions of operations and consume significant amounts of energy. Prior research has shown that a considerable number of a DNN's memory accesses and computation are redundant when performing tasks like image classification. To reduce this redundancy and thereby reduce the energy consumption of DNNs, we introduce the Modular Neural Network Tree architecture. Instead of using one large DNN for the classifier, this architecture uses multiple smaller DNNs (called modules) to progressively classify images into groups of categories based on a novel visual similarity metric. Once a group of categories is selected by a module, another module then continues to distinguish among the similar categories within the selected group. This process is repeated over multiple modules until we are left with a single category. The computation needed to distinguish dissimilar groups is avoided, thus reducing redundant operations, memory accesses, and energy. Experimental results using several image datasets reveal the effectiveness of our proposed solution to reduce memory requirements by 50% to 99%, inference time by 55% to 95%, energy consumption by 52% to 94%, and the number of operations by 15% to 99% when compared with existing DNN architectures, running on two different embedded systems: Raspberry Pi 3 and Raspberry Pi Zero. CCS Concepts: • Computing methodologies → Neural networks; Computer vision; • Computer systems organization → Embedded systems; • Hardware → Power and energy;

On the fly Deep Neural Network Optimization Control for Low-Power Computer Vision

“…Because of the importance of having adaptive workloads for different application scenarios and hardware requirements, a significant amount of research focuses on building on re-configurable DNN architectures. In these techniques, the DNNs use different layers and connections based on the input [19,20]. GaterNet [19] is a technique that uses a gating layer to deactivate certain DNN connections to reduce computation costs.…”

Low-power object counting with hierarchical neural networks

“…Hierarchical DNNs use multiple DNNs in a tree structure [13,10]. An input is characterized by the path it follows from root to leaf (Fig.…”

“…Larger DNNs are more accurate but use more resources. To obtain an acceptable tradeoff between accuracy and efficiency, we apply an architecture search technique [13]. It uses the change in accuracy density to evaluate whether a DNN (D i+1 ) with i+1 layers should be selected over a DNN (D i ) with i layers.…”

Low-Power Object Counting with Hierarchical Neural Networks

“…Our Prior Work: Our prior work [22] proposes a technique to use hierarchical DNNs for low-power image classification. To achieve high accuracy with low memory, computation, and energy requirements, the method uses the output of a DNN's softmax layer to identify and group visually similar categories.…”

“…If the value of ∆ID(D 1 , D 2 ) > T , then this process continues for increasing values of i until ∆ID(D i , D i+1 ) ≤ T . This paper selects the value of T = 0.001 by experiments [22]. A small T Fig.…”