From Shallow to Deep: Efficient Data-Driven Sensing on Mote-Scale Urban Radar and Acoustic Platforms

From Shallow to Deep: Efficient Data-Driven Sensing on Mote-Scale Urban Radar and Acoustic Platforms

Bibliographic Details
Main Author: Roy, Dhrubojyoti
Language:English
Published: The Ohio State University / OhioLINK 2020
Subjects:
Computer Science
Computer Engineering
Online Access:http://rave.ohiolink.edu/etdc/view?acc_num=osu1595409507848768
id ndltd-OhioLink-oai-etd.ohiolink.edu-osu1595409507848768
record_format oai_dc
collection NDLTD
language English
sources NDLTD
topic Computer Science
Computer Engineering
spellingShingle Computer Science
Computer Engineering
Roy, Dhrubojyoti
From Shallow to Deep: Efficient Data-Driven Sensing on Mote-Scale Urban Radar and Acoustic Platforms
author Roy, Dhrubojyoti
author_facet Roy, Dhrubojyoti
author_sort Roy, Dhrubojyoti
title From Shallow to Deep: Efficient Data-Driven Sensing on Mote-Scale Urban Radar and Acoustic Platforms
title_short From Shallow to Deep: Efficient Data-Driven Sensing on Mote-Scale Urban Radar and Acoustic Platforms
title_full From Shallow to Deep: Efficient Data-Driven Sensing on Mote-Scale Urban Radar and Acoustic Platforms
title_fullStr From Shallow to Deep: Efficient Data-Driven Sensing on Mote-Scale Urban Radar and Acoustic Platforms
title_full_unstemmed From Shallow to Deep: Efficient Data-Driven Sensing on Mote-Scale Urban Radar and Acoustic Platforms
title_sort from shallow to deep: efficient data-driven sensing on mote-scale urban radar and acoustic platforms
publisher The Ohio State University / OhioLINK
publishDate 2020
url http://rave.ohiolink.edu/etdc/view?acc_num=osu1595409507848768
work_keys_str_mv AT roydhrubojyoti fromshallowtodeepefficientdatadrivensensingonmotescaleurbanradarandacousticplatforms
_version_ 1719457608375468032
spelling ndltd-OhioLink-oai-etd.ohiolink.edu-osu15954095078487682021-08-03T07:15:53Z From Shallow to Deep: Efficient Data-Driven Sensing on Mote-Scale Urban Radar and Acoustic Platforms Roy, Dhrubojyoti Computer Science Computer Engineering Low power urban sensing on the edge of the Internet of Things (IoT) opens up a plethora of interesting and emergent machine learning applications that facilitate real-time inferencing and actuation. Concomitant with the emergence of this class of computing is an industry-wide pendulum shift that is migrating computation from the cloud to the edge, for reasons of privacy, cost, and latency. However, as these applications transition from the lab to the field, a trade-off in sensing quality and runtime efficiency needs to be navigated, since optimizing for accuracy typically produces more complex ML models with high inference duty cycle, and vice versa. It remains challenging to design models that are robust across deployment contexts, have low featurization overhead, and yet do not compromise (and ideally, improve upon) sensing quality. My research aims at achieving the joint benefits of accuracy and efficiency on a class of particularly resource-constrained IoT devices called motes, designed to last for weeks to months on battery power alone. In this thesis, I present my work in the context of two highly interesting and relevant edge applications: (a) micro-power radar classification for intruder detection in anti-poaching and surveillance contexts, (b) monitoring, analysis and mitigation of urban noise pollution through sub-wearable scale machine listening.With respect to the first application, we present one of the first truly mote-scale SVM solutions for micro-power radar sensing, designed to achieve operationally relevant discrimination when deploying in new and clutter-rich environments. To this end, we show that designing efficient features across time, frequency and amplitude domains, in conjunction with an environmentally-aware generic method for robust feature selection, can optimize performance with the fewest features over competitive solutions, thereby making it mote friendly. However, shallow SVM solutions still incur the overhead of incremental feature computation. In the next work, we introduce multi-scale, cascaded RNNs for radar sensing, and show how leveraging the ontological decomposition of a canonical classification problem into clutter vs. source classification, followed by source type discrimination on an on-demand basis can significantly improve both sensing performance as well as runtime efficiency over competitive shallow as well as deep systems. Learning these discriminators at multiple time-scales relevant to their respective tasks, and jointly training the discriminators while being cognizant of the cascading behavior between them yields the desired improvement.Urban noise discrimination in deeply embedded devices is extremely challenging because of the lack of sufficiently labeled training data, due to which state-of-the-art embedding CNNs such as Look, Listen, and Learn (L3-Net) are robustly learned from a huge pool of unlabeled datasets so that the resulting embeddings can be transfer-learned to train a variety of downstream audio classification tasks. However, with close to 4.7 million parameters, the multi-layer L3-Net CNN is prohibitively expensive for mote-scale inferencing. For the second application, we first present EdgeL3, the first L3-Net reference model compressed by 1-2 orders of magnitude for near real-time urban noise monitoring that can be run on Raspberry Pi-scale edge devices. In particular, we demonstrate the value of fine-tuning and embedding approximation in regaining the performance lost through aggressive compression strategies such as magnitude-based sparsification. Finally, we leverage coarse-grained input representation, coupled with filter-dropping and quantization to create a new CNN architecture called SONYC-L3, the first truly mote-scale embedding CNN for urban noise sensing, that outperforms state-of-the-art baselines on a large, relevant downstream dataset in spite of having >2 orders of magnitude less activation memory. 2020-11-13 English text The Ohio State University / OhioLINK http://rave.ohiolink.edu/etdc/view?acc_num=osu1595409507848768 http://rave.ohiolink.edu/etdc/view?acc_num=osu1595409507848768 unrestricted This thesis or dissertation is protected by copyright: some rights reserved. It is licensed for use under a Creative Commons license. Specific terms and permissions are available from this document's record in the OhioLINK ETD Center.

Similar Items