Adaptive Analog Feature Extraction: Algorithm and Hardware Primitives

Digitization of analog signals from wide-band phased array radars generates a large data volume. The transmission of large volume of data, often referred to as the analog data deluge, coupled with the use of many Analog-to-Digital Converters (ADC) with high sample rates increases system power. This talk will discuss the concept of adaptive analog-to-feature (AFE) extraction that directly extract low-dimensional digital features from high-dimensional analog signals. We will present a hardware platform for AFE that couples advanced signal processing and machine-learning methods with a mixed-signal Compute-In-Memory (CIM) circuit architecture. As a specific application, the talk will discuss a CIM based beamforming accelerator (BeamCIM) that uses linear embedding to transform high-dimensional analog inputs to lower dimension digital features and perform digital beamforming using those features. BeamCIM reduces both digitization (number of ADCs) and computation requirements while maintaining beamforming quality. Although, Beam-CIM shows that a significant amount of dimensionality reduction is possible for a fixed scene, it is critical, that the features adapt as the environment changes. While this concept of adaptation is standard in array processing, all previous work on this topic works from fully observed features. We are designing new techniques, using ML-based online optimization, to adapt the features as the environment changes that work directly from the “compressed” data. The talk will provide insights on the potential of adaptive AFE and discuss associated hardware design challenges.