At Rise of AI 2018 in Berlin, Daniel Kiernan, senior OEM alliances manager at NVIDIA, argues that the central challenge of AI is not a lack of data but the difficulty of turning that data into actionable insight. He walks through customer examples from healthcare, industry and customer service to show AI already at work rather than as a future technology. In healthcare, strokes kill more people worldwide than AIDS and tuberculosis combined, and Ohio State University has raised the efficiency of brain-scan diagnosis by 81 percent using an AI algorithm alongside its neurologists. In industry, GE uses sensor data and deep learning for predictive maintenance on machinery, saving more than six million dollars a year. Kiernan describes the end of Moore's Law and NVIDIA's GPU architecture, introduced around 2010, as the technology behind a predicted thousandfold performance increase by 2020, alongside rapid growth in developers and AI startups since 2013. He also explains NVIDIA's NVLink interconnect, co-developed with IBM for Power9 servers, and closes by inviting the audience to NVIDIA's GTC conference in Munich in October 2018.
Key takeaways
- Strokes kill more people globally than AIDS and tuberculosis combined, making faster diagnosis a priority AI use case, according to Kiernan.
- Ohio State University increased brain-scan diagnostic efficiency by 81 percent using an AI algorithm alongside its neurologists.
- GE saves over six million dollars a year using sensor data and deep learning for predictive maintenance instead of reactive repairs.
- NVIDIA's GPU technology, introduced around 2010, is projected to deliver a thousandfold performance increase by 2020 as Moore's Law slows.
- KLM's deep-learning chatbots raised the share of customer queries answered on social media from 1 percent to 100 percent.
Detailed notes
Healthcare: diagnosing strokes and cancer faster
Kiernan opens with a statistic he calls surprising: strokes kill more people worldwide than a number of diseases including AIDS and tuberculosis combined, and early detection is critical to survival. Ohio State University sees over 800,000 stroke cases a year in the US and has used an AI algorithm, working alongside its neurologists, to diagnose brain scans faster; Kiernan states this has increased diagnostic efficiency by 81 percent.
Predictive maintenance in industry
In manufacturing, Kiernan points to GE, which uses sensor data and video analytics with deep learning to predict machinery failures before they happen, rather than reacting once equipment breaks down. He says GE has saved over six million dollars a year by applying this preventative, deep-learning-based approach instead of dispatching technicians only after a failure.
The end of Moore's Law and the rise of GPU computing
Kiernan frames NVIDIA's GPU technology, introduced around 2010, as a response to the end of Moore's Law, the doubling of transistor counts every 18 to 24 months that began in the 1980s. He cites a prediction of a thousandfold increase in performance by 2020 compared with today, and describes rapid ecosystem growth: developer numbers have risen to over 820,000 since 2013, and the number of startups using accelerated computing and AI grew from just over 300 in 2016 to over 2,800. He notes that frameworks such as Caffe, TensorFlow and Spark are natively accelerated on NVIDIA hardware and available through cloud providers including IBM Cloud and AWS.
Automating customer service at KLM
As an example of AI at the edge, Kiernan describes KLM, which has over 215 million social media followers and employs 235 social media staff to handle customer queries. He says that before deep-learning chatbots, those staff could only answer about 1 percent of incoming queries; deep learning frameworks now poll social media and news sources to answer routine questions directly, raising the answered share from 1 percent to 100 percent, while routing more complex conversations to the human team.
NVLink and IBM server hardware
Kiernan's one technical slide covers NVLink, a chip-to-chip interconnect NVIDIA co-developed with IBM to speed up data throughput between GPUs and CPUs inside a server, and across multiple servers in a data center. On IBM's Power9 systems, he says NVLink connects NVIDIA GPUs to each other and to the CPU, with access to more than two terabytes of addressable memory, which he presents as necessary for training on very large data sets.
Quotes
Strokes actually kill more people around the world than a number of diseases such as AIDS and tuberculosis combined.
Daniel Kiernan
The Nirvana, the thing that we all want, is real-time analysis where you can get answers to your questions now.
Daniel Kiernan
Questions answered in this talk
How did Ohio State University use AI to diagnose strokes faster?
According to Daniel Kiernan at Rise of AI 2018, Ohio State University applied an AI algorithm alongside its neurologists to diagnose brain scans, increasing diagnostic efficiency by 81 percent for the roughly 800,000 stroke cases the US sees each year.
How much money did GE save using AI for predictive maintenance?
Daniel Kiernan said at Rise of AI 2018 that GE used sensor data and deep learning to predict machinery failures before they happened, saving the company over six million dollars a year.
How did KLM use deep learning to handle customer service queries?
According to Daniel Kiernan at Rise of AI 2018, KLM's 235 social media staff could only answer about 1 percent of the queries from its 215 million social media followers manually; deep learning chatbots raised that to 100 percent by handling routine questions and routing complex ones to staff.