Summary of tinyML Summit 2022: AnalogML: Analog Inferencing for System-Level Power Efficiency

00:00:00 - 00:25:00

The video discusses the capabilities of tinyML, which is very capable of doing analog pre-processing and creating features that can be fed into a neural net. Additionally, it has some mixed signal and logic capabilities that can help control timings and some extra decision making tasks. Finally, the video demonstrates how to use an analog signal processing chain to create a generic audio signal processing chain.

• 00:00:00 The speaker will discuss how analog processing can be more efficient than digital processing, and how this can be used to build low-power sensing systems. They will start by discussing how noise in the physical world can be wasted by inefficient digital processing, and go on to discuss how analog processing can be used to determine system relevance. They will conclude by talking about how analog processing can be used to build systems that can last for a long time on a battery.
• 00:05:00 The video discusses how analog ML can be more efficient than digital ML, and how the AnalogML core consists of a variety of analog blocks that are configured in an array. Software controls everything and software controls the configuration of the individual components. The analog ML array is subdivided into computational analog blocks, and each computational analog block has different analog signal processing elements and analog non-volatile memory. The analog memory is used to control the parameters of the circuits, as well as mitigate issues that typically limit the application of analog.
• 00:10:00 The video discusses the capabilities of tinyML, which is very capable of doing analog pre-processing and creating features that can be fed into a neural net. Additionally, it has some mixed signal and logic capabilities that can help control timings and some extra decision making tasks. Finally, the video demonstrates how to use an analog signal processing chain to create a generic audio signal processing chain.
• 00:15:00 The video discusses how an AnalogML-capable system can be used to power various features like spectral energy snr estimates, zero crossing rates, and classifications based on different sound events. It also discusses how the AnalogML system consumes very little power and is able to keep the system's average power consumption very low.
• 00:20:00 The speaker discusses analogML, which allows for machine learning in the analog domain. They note that this is a Pro technology, and that the activation functions are all in the analog domain.
• 00:25:00 The TinyML Summit is a yearly event where attendees discuss analogML, a technology that allows for system-level power efficiency. This year's event was sponsored by h impulse, executive sponsors were armed deep light, qualcomm incidence, platinum sponsor analog devices, brainchip infineon click attack latin ai, nxp, reality ai renaissance, sony, semiconductor, synaptics, and gold sponsors photohub micro ai prophecy, seed studio sentiment, st microelectronics, and xmos. Civil sponsors include avion devices, spinity, sierra amsa green, gravity, hi-mix hotg imagi mob, item is lattice, nota, and omni ml pixart.