At Rise of AI 2019 in Berlin, Dinakar Munagala, co-founder and CEO of Thinci, argues that four barriers are slowing AI adoption: cost, talent, time and data, and describes how new processor architectures and tooling could lower them and bring AI to the masses, much as cheap computing, operating systems and the internet drove the PC revolution and touch UIs and app stores drove the mobile revolution. He says building an AI application today needs cloud compute that gets expensive over months of training, or hardware costing up to 300,000 dollars, plus a large team of data engineers, data scientists and ML engineers, with 50 to 60 percent of time spent just collecting and cleaning data. He points to a global shortage of AI talent, citing a figure of about 300,000 AI researchers and practitioners worldwide against a McKinsey estimate of a quarter-million-person US data scientist shortfall by 2020. Munagala describes Thinci's programmable graph processor architecture, built to accelerate AI workloads more efficiently than CPUs or GPUs, and points to low-code tools, transfer learning, model marketplaces and synthetic datasets as ways to let non-experts, or "citizen data scientists," build AI applications.
Key takeaways
- Munagala frames cost, talent, time and data as the four barriers currently blocking wider AI adoption.
- He cites an estimate of only about 300,000 AI researchers and practitioners worldwide against a need for millions.
- His company Thinci built a programmable graph streaming processor as a third hardware category alongside CPUs and GPUs for AI workloads.
- He says his company spent 50 to 60 percent of its development time just collecting and cleaning data.
- He describes deep learning cutting drug-molecule screening from about a month for 20-30 molecules to two hours for roughly 10,000.
Detailed notes
The barriers slowing AI adoption
Munagala says almost every company he visits wants to be first with AI but runs into the same barriers: cost, talent, time and data. He frames AI as the latest in a sequence of historical inflection points, after the industrial, electrical, PC and mobile/internet revolutions, each of which created new industries, jobs and wealth. He argues the PC revolution needed affordable powerful computing, an operating system and good tool chains, plus the internet for sharing work, while the mobile revolution needed a natural touch UI, a personal device and an app store model. He says AI needs the same ingredients: affordable, power-efficient compute for training and edge inference, easy application building, adequate data access, pre-built models and self-customized AI for a user's specific situation.
Why compute, talent and data are so expensive today
On compute, Munagala lists three options for building an AI application: renting cloud resources, which gets expensive over a roughly six-month training project; buying a dedicated box such as a DGX for 180,000 to 300,000 dollars; or buying a cheaper GPU, which then takes about three days to train even a simple model, before facing the separate challenge of power-efficient hardware for edge deployment. On talent, he cites Tencent's estimate of only about 300,000 AI researchers and practitioners worldwide against a need for millions, and McKinsey's prediction of a US shortfall of a quarter million data scientists by 2020, arguing that today's tools are built for AI experts rather than for closing that gap. On data, he says his own company spent 50 to 60 percent of its time on getting, cleaning and labeling data before training could even start, a pattern he says holds across companies he has spoken with. He estimates the full cycle of getting data, training a model and deploying it to hardware takes months to years and costs millions of dollars.
Graph processors as a new hardware category
Munagala argues that CPUs, which exploit instruction-level parallelism, and GPUs, built for thread-level parallelism, are not well matched to the graph-shaped workloads of neural networks and CNNs. He presents this as the next step in a decades-long pattern in which new processor architectures coincide with new programming approaches, and says the industry increasingly accepts the need for a third class of processor, alongside CPUs and GPUs, suited to AI workloads for the next 20 to 30 years. Thinci's graph streaming processor is presented as this kind of programmable, AI-specific architecture, positioned to cut cost and power consumption relative to existing hardware.
Tools, marketplaces and synthetic data lowering the skills barrier
To close the skills gap, Munagala points to a shift from Python- and R-based tooling toward languages such as C++ or Java that more of the worldwide developer population already knows, plus tools that let database administrators prepare training data and let business analysts use transfer learning without being data scientists. He describes this as enabling a "citizen data scientist" who can build an application with low-code or zero-code tools. On data, he expects model marketplaces and transfer learning to grow, including last-layer training that adapts an existing model, such as one built for face recognition elsewhere, to a new setting like an office with minimal extra training. He also points to synthetic data, citing autonomous driving companies that now do an estimated 70 to 80 percent of algorithm testing in simulated environments rather than years of real-world camera driving, with the remainder still needing real-world testing.
Expected benefits once AI becomes cheap and easy
Munagala argues that if compute and tooling reach something close to PC price points, developers will spend less time on infrastructure and more on major problems. He lists agriculture, citing better crop yield prediction and matching seed types to climate risks such as typhoons; education, including AI tutors for remote areas and curricula customized to a student's strengths and weaknesses; energy and utilities, through cross-country sharing of grid-efficiency learnings; and medicine, citing a claim that deep learning has cut the time to screen molecules from about a month for 20 to 30 molecules against 150,000 protein targets to about two hours for roughly 10,000 molecules against the full human proteome. He also mentions encountering three or four startups within six months working on early cancer prediction or disease diagnosis, and expects quality healthcare for uninsured and underinsured populations and lower insurance costs as further benefits.
Questions answered in this talk
What are the main barriers to AI adoption according to Dinakar Munagala?
According to Dinakar Munagala at Rise of AI 2019, the main barriers are cost of compute, shortage of skilled talent, the long time it takes to build an AI application, and scarcity of usable data.
What is a graph streaming processor?
According to Dinakar Munagala at Rise of AI 2019, a graph streaming processor is a programmable hardware architecture, built by his company Thinci, designed as a third category alongside CPUs and GPUs to more efficiently accelerate the graph-shaped workloads of neural networks.
How much time does data preparation take in building an AI application?
According to Dinakar Munagala at Rise of AI 2019, his company found that 50 to 60 percent of development time went into getting, cleaning and labeling data before training could even begin, a pattern he says holds across other companies too.