Deep ML on GPUs: AI, industry, and R&D soar.
Time for another “thank you gamers, for fueling GPU-vendors”; that is, indirectly boosting deep-learning HW. A year ago, BioNTech and InstaDeep (a partner of London-based DeepMind) launched their AI innovation lab in London and Mainz for next-generation vaccines and biopharmaceuticals, combining InstaDeep’s DeepChain platform with BioNTech’s Covid know-how and technology. It’s already transforming research and IT-architecture alike: This week, the lab announced their high-risk Covid-variants Early Warning System (EWS), ranking the Omicron a high-risk one – on the same day its sequence became available.
This marks a “2.0” of the long-lasting win-win relationship between gene-based models and AI (see also diagram example from Informator’s course, 2020). The EWS can evaluate new variants in a matter of minutes, and it can (risk-)monitor variant lineages in real time. In other words, save months of precious time. 10,000 variant sequences being discovered every week, humans simply can’t cope.
AlphaFold2 by another InstaDeep partner in London, DeepMind, has predicted the shapes of nearly every protein in the human body, and their AI-team leader John Jumper is one of “Nature’s 10” who shaped science in 2021. Karolinska Institutet’s KI News wrote: AI may allow (…) to accurately simulate highly complex biological scenarios in silico.
(Figure by DeepMind blog: Complex 3D shapes emerge from a string of amino acids).
The IT-architect part of my brain is reasonably aware of HW-vendors who underpin this second wave of SciFi-come-true in AI apps. For 2022, InstaDeep’s HW partner NVIDIA lists Top 5 Trends to watch this year:
1. To move to production, edge AI management will become the responsibility of IT departments.
2. Multimodal AI: more intelligent applications responding to what they see, hear, sense. These employ skills like NL understanding, conversational AI, pose estimation, inspection and visualization. Expect an expansion into robotics, healthcare, and more.
3. AI and Industrial IoT: “according to Gartner, by 2027 ML & deep learning will be included in 65+ percent of edge use cases, up from less than 10 percent in 2021.” Factories can add AI onto cameras and other sensors for inspection and predictive maintenance. For instant action, RT responses are made possible by connecting the AI inference application with the IoT platforms that manage the assembly lines, robotic arms etc.
4. AI-on-5G in Industry 4.0: ultra-low latency, guaranteed quality-of-service. Expansion into plant automation, robots, monitoring/inspection, vehicle telemetry, and more.
5. ML Ops shifting to edge computing, to drive the flow to and from the edge – ingesting new data or insights from the edge, retraining, testing, and redeploying ML apps.
Finally, are we rushing into a blind alley of ever-bigger Big Data with ever-bigger hunger for resources and carbon footprint? Not necessarily. For architects, there’s most often a workaround. Not least, in innovative algorithms. Remember DeepStack.AI running on a laptop with a “gamer” NVIDIA GPU as an add-on, instead of a supercomputer?
A month ago, DeepMind’s NL processing AI RETRO (Retrieval-Enhanced Transformer) matched the performance of deep NNs that are 25 times its size. It uses an external database of text passages, like a crib, to generate new human-like sentences. Along with a drop in cost and in training time, it’s database also takes it closer to eXplainable AI. By the way, several AI researchers from Michael Bowling’s DeepStack team from Edmonton and Prague are now co-developing DeepMind’s new universal Player of Games.
Happy New 2022 to all architects, linguists, CTOs/CIOs, bioscientists, Industry 4.0 innovators, gamers, the ECDC, AI & data scientists, and dozens of other specialists!
Trainer at Informator, senior modeling and architecture consultant at Kiseldalen’s, main author: UML Extra Light (Cambridge University Press) and Growing Modular (Springer). Advanced UML2 Professional (OCUP cert level 3/3).
Milan and Informator collaborate since 1996 on architecture, AI, rules, modeling, UML, requirements, and design. You can meet him this year at these courses in English or Swedish (remote participation is offered, encouraged, and strongly recommended by epidemiology experts) :
(on demand: Modular Product Line Architecture )