Current machine learning and other AI systems rely on unsustainably large quantities of data owing to inefficient digital architectures. A shift towards upcoming intelligent analog solutions can help reduce external storage dependencies, ensuring a sustainable and sovereign AI industry.
The AI enterprise appears to be thriving with an impressive array of automated systems intended to assist and augment human living in a variety of ways from industrial robot labour, self driving cars, human language processors and home assistants to automated support and advisory services pertaining to medical, market and educational contexts.
Further, digitization in the popular awareness has come to be inextricably linked to modern information and communication technology to the extent that a term such as “Analog AI”  can seem like an oxymoron. Why then would one turn to Analog tech?
The answer lies partly in the inefficient usage and wasteful production of extremely large quantities of data associated with training, deployment and operation of conventional AI systems.
The common practice has been to train machine learning (ML) systems on ever increasing amounts of data, in the belief that the performance of such systems can be indefinitely improved by throwing ever increasing amounts of data at the model under training. However, researchers at the Allen Institute of AI  have found that while larger data sets can improve machine learning accuracy, the size of these improvements shrinks quite rapidly with extremely large data sets producing negligibly small improvements in performance.
Storage of such large data sets requires a steady reliance of external Cloud memory support maintained primarily by foreign private corporations. Further, processing these huge data sets requires a humongous amount of power which only a handful of corporations, and countries like the US and China can afford. Researchers are therefore calling for a shift away from a high resource consumption, Red AI approach to a Green AI  training approach which aims for not just high accuracy but high efficiency.
All-time systems such as home assistants and burglar alarms which are battery operated and always ON are examples of a rapidly growing industry of IoT devices. However, their digital architectures are wasteful, using up 70-90% of their batteries processing irrelevant data .For instance, a voice-activated home assistant must remain active at all times digitizing and analyzing every sound before establishing it as its “wake word” , such as ambient noises arising from traffic on streets, birds, the movement of furniture, vendors calling out their wares and the like. This vast amount of data is eventually discarded after being subjected to expensive digital processing in the Cloud.
The International Data Corporation estimates that by 2025, there will be about 41.6 billion connected IoT devices generating around 80 zettabytes (1 ZB = 10^21 bytes) of data, 80% of which will be processed unnecessarily [3,4]. Data sizes this large will then require a significant amount of Cloud support (most likely, foreign or corporate).
The solution to these unsustainably high storage dependencies lies in the use of intelligent Analog technology. Aspinity’s new RAMP chip  drastically reduces the volume of irrelevant data generated by up to a 100 times, by using frequency/energy signatures to eliminate extraneous data (eg: non-human sounds) at the analog stage and only then handing over data for higher level processing.
Aspinity’s Analog ML core provides very low resource consuming alternatives to digital home, IoT, consumer, industrial, and biomedical applications, drastically reducing the data fed to neural networks and storage requirements for sensor data. This in turn enables the extension of battery life by 10 times or more  in battery-powered smart sense applications such as face detection, predictive maintenance, behaviour monitoring, bar code recognition, leak detection, cell phone cameras, visual health monitoring, user-occupancy-based services and object tracking .
Data heavy conventional digital systems can end up being prohibitively expensive for students and researchers keen to participate in deep learning research, especially those from emerging economies. To quote Schwartz et al “Our goal is to make AI both greener and more inclusive, enabling any inspired undergraduate with a laptop to write high-quality research papers” .
Analog technology will go a long way towards ensuring that AI aspirations are not dictated by inefficient and unsustainable data dependencies and corresponding foreign support systems. A sovereign presence on the AI stage can only be established through an independent scientific disposition and avoidance of a hasty adoption of often ill-considered developmental approaches of resource-rich actors.
Data storage concerns may be addressed through elimination of both irrelevant data generation and the unprincipled usage of data. But is there a case to be made for standardization of data to obtain smaller data sets of a high quality, sharing of data to reduce redundancy (a growing practice among academia) and registered access to data libraries?
Views expressed by the author are personal and need not reflect or represent the views of Centre for Public Policy Research.