1,721,018 research outputs found
On the Reproducibility of Experiments achieved by TinyRCE
Full text linkTinyRCE is a hyperspherical classifier aimed at Continual Learning On-Tiny-Devices, a challenging task in which a Machine Learning model is required to learn from continuous streams of data while being directly installed on a (tiny) device with limited computational resources. The classifier has so far been applied to several use cases, including Human Activity Recognition, Ball Bearing Anomaly Classification, Keyword Spotting and Image Classification. The proposed work in this paper focuses on the reproducibility of TinyRCE’s experimental results already published on other papers. This to prove that all the published results are quantitatively reproducible. All the experiments have been executed on two independent computing machines to profile the impact on accuracy of the computations. As the outcomes are matching, the experimental reproducibility of TinyRCE’s accuracy over all the use cases has been positively verified
Accurate Classification of Sport Activities Under Tiny Deployability Constraints
No full textHuman Activity Recognition (HAR) plays a prominent role in various domains, such as healthcare, surveillance, and sports. In this paper, our goal is to identify the most accurate Deep Learning (DL) algorithm under tiny deployability constraints. Our results show that a Recurrent Neural Network (RNN) given by the combination of a one-dimensional Convolutional Neural Network (ID-CNN) with Bi-directional Gated Recurrent Unit (Bi-GRU) is the most attractive solution, with respect to Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and the recently introduced Legendre Memory Unit (LMU). The algorithm performance is investigated over a publicly available dataset consisting of 19 different daily activities. According to the obtained results, 1D-CNN-BiGRU has an average accuracy within 0.2% of that of BiGRU (the RNN with highest accuracy) with an execution time more than 4 times shorter
TinyRCE: Forward Learning Under Tiny Constraints
Full text linkThe challenge posed by on-tiny-devices learning targeting ultra-low power devices has recently attracted several machine learning researchers. A typical on-device model learning session processes real time streams of data produced by heterogeneous sensors. In such a context, this paper proposes TinyRCE, a forward-only learning approach based on a hyperspherical classifier aiming to be deployed on microcontrollers and, potentially, on sensors. The learning process is fed by labeled data streams to be classified by the proposed method. The classical RCE algorithm has been modified by adding a forget mechanism to discard useless neurons from the classifier's hidden layer, since they could become redundant over time. TinyRCE is fed with compact features extracted by a convolutional neural network which could be an extreme learning machine. In such case, the weights of the topology were randomly initialized instead of trained offline with backpropagation. Its weights are stored in a tiny read-only memory of 76.45KiB. The classifier required up to 40.26KiB of RAM to perform a complete on-device learning workload in 0.216s, running on an MCU clocked at 480MHz. TinyRCE has been evaluated with a new interleaved learning and testing protocol to mimic an on-tiny-device forward learning workload. It has been tested with openly available datasets representing human activity monitoring (PAMAP2, SHL) and ball-bearing anomaly detection (CWRU) case studies. Experiments have shown that TinyRCE performed competitively against a supervised convolutional topology followed by a SoftMax classifier trained with backpropagation on all these datasets
Air Quality Estimation with Embedded AI-Based Prediction Algorithms
No full textAir pollution is one of the main criticalities in cities with large populations. Therefore, accurate air quality prediction is crucial to control the environmental pollution and to maintain healthy living conditions for the citizens. To this end, particulate matters (e.g., PM 2.5) have been recognised as one of the most important pollutants with a detrimental impact on human health. In this paper, we investigate the trade-off between estimation accuracy and computational complexity of Machine Learning (ML) and Deep Learning (DL) algorithms in predicting air pollution (in terms of PM 2.5 concentration), in order to investigate their applicability to Internet of Things (IoT)-oriented applications. Six DL methods are implemented and evaluated, considering various time lags. DL approaches are shown to outperform ML approaches—in the DL case, two distinct optimizers, namely ADAM and Root Mean Squared Propagation (RMSProp), are considered. Among all algorithms evaluated, GRU had a RMSE of 20.02, while SimpleRNN reduced the MACs number by 98.90% over GRU and with an accuracy drop of 7.5%
Benchmarking MLCommons Tiny Audio Denoising with Deployability Constraints
No full textSpeech enhancement is a critical field in audio signal processing given its essentiality to overcome obstacles related to loud and damaged speech signals. Due to the revolutionary capa-bilities of Deep Learning (DL) models, there has been significant interest on benchmarking them and studying their suitability for tiny embedded systems. In this paper, we thoroughly examine the growing field of voice improvement, with a specific emphasis on the use of DL-based techniques under consideration by the MLCommons standardization. In particular, among the others, the Legendre Memory Unit (LMU) model achieves an average Scale-Invariant Signal-to-Distortion Ratio (SISDR) on 8.613 in 627 KiB of FLASH memory, making it deployable on tiny microcontrollers by requiring only 7 ms per inference run
Frame buffer-less stream processor for accurate real-time interest point detection
No full textA high performance HW accelerator is proposed to extract and refine the Interest Points from images, by accurately calculating the Difference-of-Gaussian and using refinement algorithms from the SIFT method. Unique features of the accelerator consist in an accuracy comparable to the CDVS Test Model, reference software; in the capability to process the incoming pixel in streaming order to minimize the amount of embedded memory and avoid external frame buffers; in the possibility to configure the processor with different area/speed ratios. FPGA synthesis on a Xilinx XC7V2000T returns a maximum operation frequency up of 309 MHz at the fastest corner. Standard cell synthesis with the STMICROELECTRONICS FDSOI 28 nm technology, de-congestioned by the use of DPREG memories in place of SRAM, gives a maximum frequency of 1.2 GHz and a power dissipation of about 1 W at the typical conditions
Deep Neural Quantization for Speech Detection of Parkinson Disease
No full textAmong all the diseases that nowadays people all around the world suffer, Parkinson's Disease is one of those neuro-degenerative disorders heavily impacting, and unfortu-nately expected to increase, the well-being of, especially, elderly individuals. Besides traditional medical treatments, timely and unobtrusive ways to accurately detect the onset of this disease can rely on Machine Learning (ML) and Deep Learning (DL) techniques, also because of their ability to efficiently extract information from multidimensional data on heterogeneous platforms (including, for instance, constrained Internet of Things devices). This paper presents an experimental performance evaluation of several floating point and quantized ML and DL models which can be deployed efficiently on a tiny microcontroller, namely a STM32U5 micro controller device (available in the STMicroelectronics device cloud). They have been applied to a public Italian voice speech dataset in order to classify the Parkinson Disease in three classes of patients. The experimental results demonstrate the applicability of Neural Network (NN)-based approaches for detecting the disease, as well as the deployability of traditional ML models on tiny resource-constrained devices, allowing a substantial flash memory usage reduction (when compared to non-quantized models) while keeping relatively high accuracy
Going Beyond Counting First Authors in Author Co-citation Analysis
Full text linkThe present study examines one of the fundamental aspects of author co-citation analysis (ACA) - the way co-citation
counts are defined. Co-citation counting provides the data on which all subsequent statistical analyses and mappings
are based, and we compare ACA results based on two different types of co-citation counting - the traditional type that
only counts the first one among a cited work's authors on the one hand and a non-traditional type that takes into
account the first 5 authors of a cited work on the other hand. Results indicate that the picture produced through this non-traditional author co-citation counting contains more coherent author groups and is therefore considerably clearer. However, this picture represents fewer specialties in the research field being studied than that produced through the traditional first-author co-citation counting when the same number of top-ranked authors is selected and analyzed. Reasons for these effects are discussed
- …
