Feed Your Mind - Highly-Efficient Machine Learning Without Neural Networks

#AI #ML #NeuralNetworks

Thanks to electronic enabling technologies, machine learning (ML) has gained momentum to get a pervasive paradigm in many applications due to the unquestionable results achieved. Yet, ML appears as a misused buzzword for attracting scientific interest and curiosity. The main problem of ML is the vastity of the tools and techniques mostly linked to neural network architec-tures. In this presentation, I will show how a niche of ML, based on statistical inference, could achieve highly reliable predictive models on reduced da-tasets. Instead of increasing the architectures to follow the algorithmic com-plexity, I will show a reverse approach, where the simplicity of the predictive models could be embedded in simple, low-power architectures. The main ad-vantages of these techniques are first to avoid complex elaborations, being suitable for edge computing, and secondly to develop tools on a robust mathe-matical background, such as multivariate analysis helping the developers to a critical understanding of the raw data. I will show results on different appli-cations based on spectral sensing information processing in precision agricul-ture and structural engineering and how this approach could be extended to any kind of spectral sensing (microwave, mm-wave, terahertz radiation, ultra-sound, optical, infra-red, mass spectroscopy).