• artificial intelligence
• FPGA
• HW-SW codesign
• on-the-edge applications
• onboard application
• satellite systems

Artificial Intelligence was born in 1956 as a discipline of Computer Science, aiming to simulate human intelligence via machine algorithms.
Its history sees up and down moments which ended with “AI winter”, usually due to the hardware limitations or lack of practical solutions.
In the 1990s, with the arrival of new metal-oxide-semiconductor (MOS) and complementary-MOS (CMOS) transistors, AI took new life.
This technological advancement has encouraged AI algorithms adoption in different sectors, from assistive technologies to space.

In the late 2010s, AI was completely reborn, powered by some new applications, such as DeepBlu, able to defeat a man in a match of GO, considered one of the most difficult existent games.
However, these incredible applications able to learn directly from the data have been relegated to the big server-farms and exploited using the cloud computing paradigm.
This thesis analyses pros and cons of the existing machine learning hardware accelerators starting from cloud computing with GPUs to edge accelerators also including proof-of-concept demonstrator for case studies in the field of assistive technologies and space applications.
In particular, it starts by analysing the General Purposes-GPUs, the most powerful Single Instruction Multiple Data architecture. This technology has been evaluated in the framework of the Parloma project, which aims to simplify the interaction between deafblind people by using a dedicated neural network run on a GPU, which extracts human body's joints to send to a robotic arm with a sensorised skin.

Although very flexible, the cloud computing powered by GPU accelerators presents some limitations, such as data transportation, security and, migration.
The edge paradigm is trying to overcome these problems, integrating the processing platforms where the data are acquired.
To this end, the most interesting and innovative edge accelerators were studied, characterising both on hardware and software sides' strengths and weaknesses of those solutions.
Then, the above-mentioned characteristics have been demonstrated through three different use cases.
Rimedio is a robotic arm mounted on a power wheelchair and controlled through a dedicated smart human-machine interface. It is able to detect and highlight only the object of interest using a YOLOv2 neural network.
The entire robot control system runs on a Raspberry Pi 3, while the neural network was tested both on GPU and Myriad 2 VPU, demonstrating several advantages of the on the edge approach.

AI-Drive uses a more powerful accelerator, the Nvidia Jetson Nano capable to run multiple networks at the same time, thanks to its GPU-like architecture.
The project aims to assist people who are unable to react quickly to possibly dangerous situations to drive a power wheelchair, exploiting two different neural networks for detecting and avoiding obstacles.

Furthermore, pushed by the hype generated by the AI fever of the latest years, a COTS hardware accelerator for machine learning and computer vision applications was validated and characterised for satellite on-board applications, considering the additional requirements deriving the space environment.
The Phi-Sat-1 mission represents the first technological demonstrator of a deep neural network, called CloudScout, executed directly on-board satellite.
The CloudScout network is a small segmentation network, which classifies each pixel of a hyperspectral image taken as input in two classes: cloudy and not cloudy.
The output map provided by the network is then post-processed to compute the percentage of cloudy pixels contained in each image.
This algorithm represents a great revolution, since the computation of the map directly on-board allows to reduce the transmission bandwidth, strongly limited in small- and nano-satellites, focusing only on images that contain less than 70% of cloudy pixels.
Finally, analysing the limitations of the embedded accelerators for these applications, a low-power GP-GPU for FPGA, inside the ICU4SAT project, is given, trying to provide a stronger support to the innovative solutions announced in the latest years, while reducing programmers’ effort.