While an exponential growth in data and increase in compute power have been key drivers of AI innovation, without research — and the tremendous effort academics and companies invest in it — these advances wouldn’t go very far.
At his Transform 2019 talk last month, Lyft’s head of product for machine learning, Gil Arditi, turned a spotlight on this research. He ripped through example after example of research that has propelled head-turning advances in AI and machine learning in the past year. Below are just a few of the examples he covered (watch his entire talk above).
Deep reinforcement learning
Two AI companies that have been at the center of deep reinforcement learning have used video games as their research playground. OpenAI Five has used the technology to defeat humans in multiplayer online battle arena video game Dota 2 — and has won 4,075 games against human players for a victory rate of 99.4%. Alphabet’s DeepMind is demonstrating similar success pitting its AlphaStar AI agent against real gamers in Starcraft II. During a livestream against pros, AlphaStar swept all 10 matches.
Arditi noted that these accomplishments, while impressive, took considerable effort. “OpenAI Five trained for more than 45,000 years of gameplay, and AlphaStar took more than $26 million in the monetized value of the compute resources they used.”
Natural language processing
Until recently, Google-developed BERT solved a wide range of problems in NLP, including things like sentence classification, sentence space similarity, questions, and answers. Arditi asserts that XLnet, which emerged from a research paper several weeks ago, is an up-and-coming contender. “It’s outperforming BERT on some new benchmarks,” he explained.
GPT-2 model for text prediction
You can think of GPT-2 as the world’s best text generator — if the model is given a starter sentence, it generates subsequent sentences in the same theme.
Arditi demonstrated with the input: “A train carriage containing controlled nuclear materials was stolen in Cincinnati today. Its whereabouts are unknown.” GPT-2 continued, “The incident occurred on the downtown train line, which runs from Covington and Ashland stations. In an email to Ohio news outlets, the U.S. department of energy said it is working with the federal railroad administration to find the thief.” As Arditi noted, “It’s a huge step forward that also creates some potential for fake news. This is very timely, because of the upcoming election cycle.”
Identifying fake news
Of course, every problem technology creates requires a new solution. “Fake news is something that’s very hard for humans to understand [and] to see,” said Arditi. Enter Grover, which is being touted as a state-of-the-art defense against fake news. “This algorithm is doing really well in both generating text and in classifying text as being fake, as being not real, not generated by humans. It’s also looking at the source. It can adapt to the style of a paper like the Washington Post versus a lesser-known blog.”
Arditi also cited generative models, the generation of high-res images and video (which will have significant impact on both video game development and sophisticated productions that currently require armies of animators), style transfer in 3D, and more.