scholarly journals A Framework for Ultra Low-Power Hardware Accelerators Using NNs for Embedded Time Series Classification

2021 ◽  
Vol 12 (1) ◽  
pp. 2
Author(s):  
Daniel Reiser ◽  
Peter Reichel ◽  
Stefan Pechmann ◽  
Maen Mallah ◽  
Maximilian Oppelt ◽  
...  
Keyword(s):  
Time Series ◽  
Embedded System ◽  
Low Power ◽  
Holistic Approach ◽  
System Optimization ◽  
Time Series Classification ◽  
Ultra Low Power ◽  
Classification Tasks ◽  
Definition Of ◽  
Process Detection

In embedded applications that use neural networks (NNs) for classification tasks, it is important to not only minimize the power consumption of the NN calculation, but of the whole system. Optimization approaches for individual parts exist, such as quantization of the NN or analog calculation of arithmetic operations. However, there is no holistic approach for a complete embedded system design that is generic enough in the design process to be used for different applications, but specific in the hardware implementation to waste no energy for a given application. Therefore, we present a novel framework that allows an end-to-end ASIC implementation of a low-power hardware for time series classification using NNs. This includes a neural architecture search (NAS), which optimizes the NN configuration for accuracy and energy efficiency at the same time. This optimization targets a custom designed hardware architecture that is derived from the key properties of time series classification tasks. Additionally, a hardware generation tool is used that creates a complete system from the definition of the NN. This system uses local multi-level RRAM memory as weight and bias storage to avoid external memory access. Exploiting the non-volatility of these devices, such a system can use a power-down mode to save significant energy during the data acquisition process. Detection of atrial fibrillation (AFib) in electrocardiogram (ECG) data is used as an example for evaluation of the framework. It is shown that a reduction of more than 95% of the energy consumption compared to state-of-the-art solutions is achieved.

scholarly journals FilterNet: A Many-to-Many Deep Learning Architecture for Time Series Classification

2020 ◽  
Author(s):  
Robert D. Chambers ◽  
Nathanael C. Yoder
Keyword(s):  
Time Series ◽  
Deep Learning ◽  
Activity Recognition ◽  
Sensor Data ◽  
Time Series Classification ◽  
Discrete Events ◽  
Segmentation Accuracy ◽  
Model Size ◽  
Classification Tasks

We present and benchmark FilterNet, a flexible deep learning architecture for time series classification tasks, such as activity recognition via multichannel sensor data. It adapts popular CNN and CNN-LSTM motifs which have excelled in activity recognition benchmarks, implementing them in a many-to-many architecture to markedly improve frame-by-frame accuracy, event segmentation accuracy, model size, and computational efficiency. We propose several model variants, evaluate them alongside other published models using the Opportunity benchmark dataset, demonstrate the effect of model ensembling and of altering key parameters, and quantify the quality of the models’ segmentation of discrete events. We also offer recommendations for use and suggest potential model extensions. FilterNet advances the state of the art in all measured accuracy and speed metrics on the benchmarked dataset, and it can be extensively customized for other applications.

scholarly journals FilterNet: A Many-to-Many Deep Learning Architecture for Time Series Classification

2020 ◽  
Author(s):  
Robert D. Chambers ◽  
Nathanael C. Yoder
Keyword(s):  
Time Series ◽  
Deep Learning ◽  
Activity Recognition ◽  
Sensor Data ◽  
Time Series Classification ◽  
Discrete Events ◽  
Segmentation Accuracy ◽  
Model Size ◽  
Classification Tasks

We present and benchmark FilterNet, a flexible deep learning architecture for time series classification tasks, such as activity recognition via multichannel sensor data. It adapts popular CNN and CNN-LSTM motifs which have excelled in activity recognition benchmarks, implementing them in a many-to-many architecture to markedly improve frame-by-frame accuracy, event segmentation accuracy, model size, and computational efficiency. We propose several model variants, evaluate them alongside other published models using the Opportunity benchmark dataset, demonstrate the effect of model ensembling and of altering key parameters, and quantify the quality of the models’ segmentation of discrete events. We also offer recommendations for use and suggest potential model extensions. FilterNet advances the state of the art in all measured accuracy and speed metrics on the benchmarked dataset, and it can be extensively customized for other applications.

scholarly journals AN APPROACH ADOPTED FOR SMART DATA GENERATION AND VISUALIZATION PROBLEMS

2020 ◽  
Vol VI-4/W2-2020 ◽  
pp. 41-45
Author(s):  
C. Capodiferro ◽  
M. Mazzei
Keyword(s):  
Time Series ◽  
Low Power ◽  
Data Management ◽  
Data Management System ◽  
Area Network ◽  
Data Generation ◽  
Ultra Low Power ◽  
Management Platform ◽  
Sensor Devices ◽  
Smart Data

Abstract. This paper discusses how to address and overcome some of the problems related to smart data generation and visualization, such as the poor autonomy of wireless sensor devices and the flexibility of the data management platform. We described the implementation and field experiment of a modular IoT application for Smart-Farming, in which the sensor devices are powered by an on-board battery and the data management system is based on a highly flexible software stack, capable of displaying time series graphs and processing millions of data per second. The experiment shown that the power consumption of the sensor devices depends on many factors and that the lifecycle of the devices can reach years using ultra-low power processors, low power wide area network (LPWAN) such as LoRaWAN and a mix of energy saving techniques. The data management and visualization platform shown to be able to display many types of time-series graphs, deal with a wide variety of data-sources and effectively manage a large amount of data.

scholarly journals Genetic Programming and Serial Processing for Time Series Classification

2014 ◽  
Vol 22 (2) ◽  
pp. 265-285 ◽  
Author(s):  
Eva Alfaro-Cid ◽  
Ken Sharman ◽  
Anna I. Esparcia-Alcázar
Keyword(s):  
Time Series ◽  
Genetic Programming ◽  
Time Series Data ◽  
Series Data ◽  
Time Series Classification ◽  
Classification Problems ◽  
Serial Processing ◽  
Processing Approach ◽  
Very Large Datasets ◽  
Classification Tasks

This work describes an approach devised by the authors for time series classification. In our approach genetic programming is used in combination with a serial processing of data, where the last output is the result of the classification. The use of genetic programming for classification, although still a field where more research in needed, is not new. However, the application of genetic programming to classification tasks is normally done by considering the input data as a feature vector. That is, to the best of our knowledge, there are not examples in the genetic programming literature of approaches where the time series data are processed serially and the last output is considered as the classification result. The serial processing approach presented here fills a gap in the existing literature. This approach was tested in three different problems. Two of them are real world problems whose data were gathered for online or conference competitions. As there are published results of these two problems this gives us the chance to compare the performance of our approach against top performing methods. The serial processing of data in combination with genetic programming obtained competitive results in both competitions, showing its potential for solving time series classification problems. The main advantage of our serial processing approach is that it can easily handle very large datasets.

scholarly journals Regularizing Fully Convolutional Networks for Time Series Classification by Decorrelating Filters

2019 ◽  
Vol 33 ◽  
pp. 10003-10004 ◽  
Author(s):  
Kaushal Paneri ◽  
Vishnu TV ◽  
Pankaj Malhotra ◽  
Lovekesh Vig ◽  
Gautam Shroff
Keyword(s):  
Neural Networks ◽  
Time Series ◽  
Deep Neural Networks ◽  
Training Data ◽  
Time Series Classification ◽  
Strong Correlations ◽  
Average Correlation ◽  
Convolutional Networks ◽  
Fully Convolutional Networks ◽  
Classification Tasks

Deep neural networks are prone to overfitting, especially in small training data regimes. Often, these networks are overparameterized and the resulting learned weights tend to have strong correlations. However, convolutional networks in general, and fully convolution neural networks (FCNs) in particular, have been shown to be relatively parameter efficient, and have recently been successfully applied to time series classification tasks. In this paper, we investigate the application of different regularizers on the correlation between the learned convolutional filters in FCNs using Batch Normalization (BN) as a regularizer for time series classification (TSC) tasks. Results demonstrate that despite orthogonal initialization of the filters, the average correlation across filters (especially for filters in higher layers) tends to increase as training proceeds, indicating redundancy of filters. To mitigate this redundancy, we propose a strong regularizer, using simple yet effective filter decorrelation. Our proposed method yields significant gains in classification accuracy for 44 diverse time series datasets from the UCR TSC benchmark repository.

TCN Mapping Optimization for Ultra-Low Power Time-Series Edge Inference

2021 ◽  
Author(s):  
Alessio Burrello ◽  
Alberto Dequino ◽  
Daniele Jahier Pagliari ◽  
Francesco Conti ◽  
Marcello Zanghieri ◽  
...  
Keyword(s):  
Time Series ◽  
Low Power ◽  
Ultra Low Power
Sign in / Sign up

Export Citation Format

Share Document