Search

In the context of TinyML, many research efforts have been devoted to designing forward topologies to support On-Device Learning. Reaching this target would bring numerous advantages, including reductions in latency and computational complexity, stronger privacy, data safety and robustness to adversarial attacks, higher resilience against concept drift, etc. However, On-Device Learning on resource constrained devices poses severe limitations to computational power and memory. Therefore, deploying Neural Networks on tiny devices appears to be prohibitive, since their backpropagation-based training is too memory demanding for their embedded assets. Using Extreme Learning Machines based on Convolutional Neural Networks might be feasible and very convenient, especially for Feature Extraction tasks. However, it requires searching for a randomly initialized topology that achieves results as good as those achieved by the backpropagated model. This work proposes a novel approach for automatically composing an Extreme Convolutional Feature Extractor, based on Neural Architecture Search and Bayesian Optimization. It was applied to the CIFAR-10 and MNIST datasets for evaluation. Two search spaces have been defined, as well as a search strategy that has been tested with two surrogate models, Gaussian Process and Random Forest. A performance estimation strategy was defined, keeping the feature set computed by the MLCommons-Tiny benchmark ResNet as a reference model. In as few as 1200 search iterations, the proposed strategy was able to achieve a topology whose extracted features scored a mean square error equal to 0.64 compared to the reference set. Further improvements are required, with a target of at least one order of magnitude decrease in mean square error for improved classification accuracy. The code is made available via GitHub to allow for the reproducibility of the results reported in this paper.

While many companies are currently leveraging on Cloud, data centres and specialized hardware (e.g., GPUs and TPUs) to train very accurate Machine Learning models, the need to deploy and run these models on tiny devices is emerging as the most relevant challenge, with a massive untapped market. Although Automated Machine Learning and Neural Architecture Search frameworks are successfully used to find accurate models by trying a small number of alternatives, they are typically performed on large computational platforms and they cannot directly deal with deployability, leading to an accurate model which could result undeployable on a tiny device. To bridge the gap between these two worlds, we present an approach extending these frameworks to include the constraints related to the limited hardware resources of the tiny device which the trained model has to run on. Experimental results on two benchmark classification tasks and two microcontrollers prove that our AutoTinyML framework can efficiently identify models which are both accurate and deployable, in case accepting a reasonable reduction in accuracy compared to a significant reduction in hardware usages, without applying any quantization techniques of the model.

Deep Neural Networks are increasingly deployed on tiny systems such as microcontrollers or embedded systems. Notwithstanding the recent success of Deep Learning, also enabled by the availability of Automated Machine Learning and Neural Architecture Search solutions, the computational requirements of the optimization of the structure and the hyperparameters of Deep Neural Networks usually far exceed what is available on tiny systems. Therefore, the deployability becomes critical when the learned model must be deployed on a tiny system. To overcome this critical issue, we propose a framework, based on Bayesian Optimization, to optimize the hyperparameters of a Deep Neural Network by dealing with black-box deployability constraints. Encouraging results obtained on a classification benchmark problem on a real microcontroller by STMicroelectronics are presented.

This paper presents the complete design, the hardware implementation, and the experimental validation of a distributed intelligence Internet-of-Things (IoT) system, based on Heterogeneous Event-Driven Edge Mesh Architecture (HED-EMA). The latter differentiates from most common computational schemes (i.e. Cloud computing) since it distributes intelligence at local level. The proposed IoT system has a bottom-up structure composed by three main layers: Sensing/Actuation, Edge, and Cloud. It exploits commercial devices, specifically programmed and configured in order to operate within the IoT paradigm. Aiming to validate the proposed system, experimental data coming from two specific setups (home-setup and automotive-setup) will be presented and discussed. Data are extracted by a distributed sensors network, real-time collected, and processed. The lower response time compared to standard Cloud architectures and the whole system power consumption reduction are then demonstrated.

This paper presents the development of an Internet-of-Things (IoT) Cooperative System (IoT-CS) based on local Event-Driven response. After introducing IoT basic concepts and most common computational schemes (with particular attention on the differences between Cloud, Fog, and Cooperative computing paradigms), an innovative scheme for IoT will be presented. Thanks to the introduction and development of several intermediate layers between the sensor network and the Cloud, the hereby proposed system allows overcoming most of the problems of the state-of-the-art paradigms. In order to validate the proposed solution the most relevant aspects of this implementations will be presented.

We propose a novel modular CNN architecture that provides semantic segmentation and understanding of outdoor street environment images. Our solution processes a 512x1024 resolution image on a single Titan Xp GPU at 37.4 FPS attaining 70.4% IoU on the Cityscapes test dataset.

Accurate pedestrian detection has an important role in automotive applications because, by issuing warnings to the driver and acting actively on the car brakes, it can save human lives and decrease the probability of injuries. In order to achieve adequate accuracy, detectors require training sets containing a very large number of negative samples, which can be challenging for the training algorithms of models like support vector machines (SVM). A common approach to deal with such large datasets is Hard Negative Mining (HNM), which avoids working on the full set by growing an active pool of mined samples. A more recent method is the Block-Circulant Decomposition, which achieves the accuracy of HNM at a lower computational cost by reformulating the problem in the Fourier domain. The method however results in additional memory, required during training by the FFT transform, which could be reduced significantly by using only the positive examples. To address the problem, this paper proposes two main contributions: (1) it shows that the circulant decomposition method works with the same performances when only the positive samples are used in the training phase (2) it compares the performance of a detection pipeline based on HOG features trained with either both all negative and positive samples or with only positive samples on the INRIA pedestrian dataset.

Local Visual detectors and descriptors have been studied for many years, but their applications, for example on visual search, for large volume, low cost, embedded systems have been limited or negligible to date. One reason is the lack of an industry standard applicable at worldwide level. MPEG Compact Descriptors for Visual Search (CDVS) working group is filling the gap by defining a highly performing extraction stage and the bitstream syntax at its output. As a contribution to such process, this work reviews existing methods for interest point detection and description. They are compared on several databases, exploiting the pairwise matching test environment devised by the MPEG CDVS standardization group. © 2013 IEEE.