Binary neural networks at the edge

In the last decade, Artificial Intelligence demonstrated to achieve great advancements in many areas such as (but not limited to) Computer Vision, Natural Language Processing, Reinforcement Learning, and Autonomous Driving. Nowadays, AI systems potentially overcome the intelligence observed in human beings as the ambition of developing AI models with cognitive, learning, and problem-solving abilities is the driving force in AI research and development. Achieving this goal would imply the creation of highly sophisticated AI systems capable of generalizing knowledge, learning from diverse data sources, and exhibiting a level of adaptability and creativity comparable to human intelligence. Most of the performance improvements obtained by AI systems have been reached by increasing the complexity and memory footprint of the models, leading to solutions whose deployment is most of the time prevented on resource-constrained hardware. Even though many works in literature addressed the problem of compressing and reducing the computational overhead of a neural network, a lot of effort is still required to successfully deploy a complex AI solution on tiny/embedded devices. Additionally, current AI solutions achieve poor performances at adapting incrementally to new data. This phenomenon, named catastrophic forgetting, happens naturally in Deep Learning architectures when a classical training algorithm, such as backpropagation, is applied. By learning through a sequence of experiences incrementally, where the model cannot access old training data, the model knowledge falls resulting in the forgetting of past samples.