Description
Data and signal processing in Internet of Things devices requires extremely low-power operation. This can be obtained on one hand by using low-power hardware, and on the other hand by exploiting algorithms and implementations that reduce power consumption to a minimum. One example of such approaces is the Compressed Sensing, that is a recently introduced paradigm for the simultaneous acquisition and compression of an input signal. Also deep neural networks, a very commonly adopted solution for solving many tasks, may benefit in terms of energy from advanced techniques such as quantization-aware training, pruning, and the definition of new map-reduce paradigms. This research activity studies various aspects of advanced low-power data processing, including: compressed sensing acquisition; ultra-low-power Analog-to-Information converters; in-memory computing; reducing hardware complexity for efficient Deep Neural Network Implementation; hardware-algorithm co-design for data streaming processing in low-energy IoT nodes (e.g. efficient streaming PCA).