Deep Task-Based Analog-to-Digital Conversion

Analog-to-digital converters (ADCs) allow physical signals to be processed using digital hardware. Their conversion consists of two stages: Sampling, which maps a continuous-time signal into discrete-time, and quantization, i.e., representing the continuous-amplitude quantities finite number bits. ADCs typically implement generic uniform mappings that are ignorant task for is acquired, can costly when operating in high rates fine resolutions. In this work we design task-oriented learn from data how map an analog representation such system efficiently carried out. We propose model sampling quantization facilitates learning non-uniform data. Based on learnable ADC mapping, present mechanism optimizing hybrid acquisition comprised combining, tunable with fixed rates, processing, by jointly its components end-to-end. Then, show one exploit systems as deep networks optimize given utilizing Bayesian meta-learning techniques. evaluate proposed task-based case studies: first considers synthetic multi-variate symbol detection, where multiple simultaneously acquired order recover set discrete symbols. The second application beamforming channel ultrasound imaging. Our numerical results demonstrate approach achieves performance comparable resolution while reduced overall bit rate. For instance, enable accurate reconstruction images $12.5\%$ bits used conventional ADCs.