Tiny Machine Learning Environment: Enabling Intelligence on Constrained Devices

Running machine learning algorithms (ML) on constrained devices at the extreme edge of the network is problematic due to the computational overhead of ML algorithms, available resources on the embedded platform, and application budget (i.e., real-time requirements, power constraints, etc.). This required the development of specific solutions and development tools for what is now referred to as TinyML. In this dissertation, we focus on improving the deployment and performance of TinyML applications, taking into consideration the aforementioned challenges, especially memory requirements. This dissertation contributed to the construction of the Edge Learning Machine environment (ELM), a platform-independent open-source framework that provides three main TinyML services, namely shallow ML, self-supervised ML, and binary deep learning on constrained devices. In this context, this work includes the following steps, which are reflected in the thesis structure. First, we present the performance analysis of state-of-the-art shallow ML algorithms including dense neural networks, implemented on mainstream microcontrollers. The comprehensive analysis in terms of algorithms, hardware platforms, datasets, preprocessing techniques, and configurations shows similar performance results compared to a desktop machine and highlights the impact of these factors on overall performance. Second, despite the assumption that TinyML only permits models inference provided by the scarcity of resources, we have gone a step further and enabled self-supervised on-device training on microcontrollers and tiny IoT devices by developing the Autonomous Edge Pipeline (AEP) system. AEP achieves comparable accuracy compared to the typical TinyML paradigm, i.e., models trained on resource-abundant devices and then deployed on microcontrollers. Next, we present the development of a memory allocation strategy for convolutional neural networks (CNNs) layers, that optimizes memory requirements. This approach reduces the memory footprint without affecting accuracy nor latency. Moreover, e-skin systems share the main requirements of the TinyML fields: enabling intelligence with low memory, low power consumption, and low latency. Therefore, we designed an efficient Tiny CNN architecture for e-skin applications. The architecture leverages the memory allocation strategy presented earlier and provides better performance than existing solutions. A major contribution of the thesis is given by CBin-NN, a library of functions for implementing extremely efficient binary neural networks on constrained devices. The library outperforms state of the art NN deployment solutions by drastically reducing memory footprint and inference latency. All the solutions proposed in this thesis have been implemented on representative devices and tested in relevant applications, of which results are reported and discussed. The ELM framework is open source, and this work is clearly becoming a useful, versatile toolkit for the IoT and TinyML research and development community.