Deep learning was the most dominant topic of the conference, for which there were several industry-specific applications. For example, Pinterest has begun moving from gradient boosting decision trees (GBDTs) to deep neural nets for predicting various dimensions of user behavior.
In healthcare, image recognition may make a big impact sooner than many other applications. Highlighting this and exemplifying the democratization of AI, we heard a fireside chat with 17-year-old Abu Qader, who built a system to identify breast cancer tumors using publicly available datasets and Google’s open source deep learning toolkit, TensorFlow. AI is even applied to early diagnosis and prevention of diseases like heart disease, cancer and type 2 diabetes. With all the hype about the dangers of AI, it’s refreshing to acknowledge the life-saving benefits of this technology!
Andrew Ng, one of the most well-known machine learning researchers in the world, not to mention Stanford professor, Coursera founder, Google Brain founder, and former Chief Scientist for Baidu, gave a great overview of the state of AI today, wherein almost all applications use supervised learning. He outlined the strategy that he believes companies must follow to be successful at leveraging AI.
Unsurprisingly, this strategy is all about the data, which provides the fuel for the AI. This includes a focus on centralizing data to make it easier to access, and acquiring proprietary data. With many of the world’s cutting-edge algorithms being released as open source software, today is it this proprietary data, rather than other forms of intellectual property (e.g. algorithms or other technology) that will provide the competitive advantage to AI companies.
On a contrary note to all the deep learning excitement and hype, Stephen Merity of Salesforce Research (formerly MetaMind) gave an interesting talk about the limits of deep learning. He argued that deep learning is a “jackhammer” and not always the right tool for the job as it requires large volumes of training data, long train times, and expensive hardware. He suggested that it can be better to iterate on simpler or traditional algorithms, until the trade-off for the expense of deep learning can be justified.
Several presenters hit on the problem of ethical considerations and bias in machine learning. A talk by two attorneys – Daniel Guillory (Autodesk) and Matthew Scherer (Littler Mendelson, PC) – discussed a variety of ways bias could lead to profound consequence, as well as some ways to address and combat it.
With so many venues to choose from, it was impossible to attend them all. There were a range of companies, both startups and those offering ML platforms. There were AI researchers in the transportation industry that spoke on “how affordable and reliable sensors enable computer-vision-based autonomous driving and how to train vision models for object detection.” Other AI/deep learning applications included how Instacart uses deep learning to optimize the in-store shopping experience and a preview of Intuit’s AI driven self-filing tax system. You can get access to all of the conference sessions on the O’Reilly website.
If you want to learn more about my presentation subject matter – conversational AI – contact us at Nuance.