I’m interested in applying neuromorphic computing to industrial ultrasonic test data streams, among other applications. The main idea is to determine if a drastic improvement in anomaly detection can be achieved via a kT-RAM approach. Ultrasound testing of steel cylinders has become mainstream in the compressed gas industry, though all current commercial testing systems are based on classical computing architectures, and to my knowledge most are Intel / Windows based systems. My sense is that the current architectures and algorithms are grossly inefficient compared to a neuromorphic, memrister based approach, and that major improvements in speed-of-processing and accuracy of anomaly detection is possible. If this compressed gas cylinder case could be proven, then a much broader application of industrial ultrasound testing would become feasible. The goal is to have near-real-time visualization of streaming ultrasound data streams coming in from test equipment (piezoelectric transducers), which would be akin to “looking” through the steel plates as if they were transparent, with anomalies plainly visible. This could revolutionize industrial testing for bridges, buildings, pipelines, aircraft, etc. Also, if the fundamentals of neuromorphic industrial ultrasound testing could be established, then those same principles could be applied to the more challenging area of human medical ultrasound, which, if similar major improvements in image speed and resolution were achieved, an enormous advance in healthcare could be realized. Please advise if you think this would be a reasonably practical area of investigation with your new computing technology.
Anomaly detection is one of the many machine learning application types that we have proven to work on top of the Knowm tech stack. In fact an app we call “knowm Anomaly” is the first commercially available app that we are selling using the Knowm SDK. We absolutely recognize that anomaly detection is a huge market, and we’re really excited. Additionally, we will be creating apps in robotics, classification, vision, NLP, combinatorial optimization, clustering, and more.
So yes, we absolutely think that “[anomaly detection] would be a reasonably practical area of investigation with your new computing technology.” In the short term, we’ve made a generic real-time horizontally scalable anomaly detection app (Knowm Anomaly), which can be quickly deployed and tapped into any data streams quite easily. In the long term (kT-RAM), which is the very exciting part, the improvements in speed, power consumption and portability will be 5-10 orders of magnitude and close to or matching biological efficiency. The neat thing is that the people developing apps right now on top of the Knowm stack will later down the road be able to replace the emulators with real kT-RAM chips without needing to change anything else. People selling those apps will have a very clear advantage over their competition.
We offer via our Knowm Development Community a mechanism to work with individuals and companies who specialize in one area to build apps and even low level libraries within and on top of the Knowm tech stack and share in future profits. If you are interested, send us a proposal. We’re excited to work with others and push this technology forward together.