Alexy Khrabrov doesn’t just want to tell people about AI. He wants to show you, immerse you and get you as excited as he is. The founder and CEO of By the Bay and Chief Scientist at the Cicero Institute has made a career out of not only understanding and developing AI, but bringing startups and enterprises into the fold, as well, with his developer meetups and conferences through AI By the Bay.
AI By The Bay’s latest conference happen March 6 through 8 in San Francisco, and Khrabrov has made it an event that developers are clamoring to attend. Attendees don’t just hear about AI -- they get a chance to see live coding of major open-source AI and hear from some of the leaders in the field about their newest innovations.
Khrabrov was kind enough to not only give me his take on the future of AI, but to tell me more about his upcoming conference.
Tell me about yourself.
I’m currently the Chief Scientist of Cicero.AI, a community-based research institute. I run seven different meetups in San Francisco on data engineering and data science. I advise startups in the space and bring together the best developers, data scientists and founders who build startups. Previously, I was Chief Scientist at an established company, a director of analytics, a software engineer myself, and I build communities around technologies I’m passionate about.
Obviously AI is hot in many different spaces, from marketing to self-driving cars. Why is it so hot now when it’s been around for a long time, and what do you see as the key impact today?
Neural networks have been around since the 1980s -- my academic advisor at Dartmouth was an editor of a neural network magazine that was popular in the 1980s. Then in the 1990s, Support Vector Machines emerged, a technique that performed better, and people thought that neural networks were dead. But essentially, the reason people thought neural networks were dead was because they did not have enough data.
By now we know that all the advances in algorithmic deep learning were mostly known earlier. Andrew Ng has a famous talk about why deep learning got so big so suddenly, where he draws a big rocket with a huge engine and a huge fuel tank. We now have this giant fuel tank, which is the data, and we also have this giant engine, which is the computing power. Before, we couldn’t put them together, but now we can.
It turns out that the trick we needed to employ for deep learning work is to have more layers, where previously we used only a single layer or just a few layers, because the human brain has multiple layers. It’s really magical: If you throw in a bunch of layers, you get a powerful representation of the world. Then you throw a bunch of data in which became available through the traffic of the Internet, devices, smartphones and cloud computing, and you get results. Everything comes together, creating a moment in time when all these techniques can fire. I think we are in this moment. There is no revolution or breakthrough in machine learning. It’s a confluence of capabilities which makes it all work.
Let’s talk about the three elements of your event.
We have a three-day conference, meant for technical audiences, meetups, and founders where we can de-hype AI. I want to structure the conference as a deep dive and overview of AI.
The first day is called Oh.Hai.AI, a live coding overview of different techniques. Instead of people just talking about the stuff, I want to see AI in action. We open with Salesforce's Einstein -- the VP of Data Engineering and Data Science, Vitaly Gordon -- who will actually come on stage and show the guts of Salesforce science, which is Scala and Spark, tools which can crunch consumer data at web scale. He is going to show how machine learning is implementing AI, which customers will see. We have, for instance, talks about attention and memory in deep learning networks.
We’ll have talks that will teach people what pieces of AI are working already and try to “show the money.” This is the major goal of the day, that moment in a technical talk when people switch from introductory bullet points to the actual code, and you can almost hear the room going silent. The code shows what really works, what really exists, and it changes the tone of the conversation.
The second day is called Self.Driving.Cars. We don’t have them just because they are hot; self-driving cars are a very well-defined problem. If you hear people talk about the AI business, you hear a lot of buzzwords like “synergy” and “adding value,” which are also the buzzwords mentioned in the context of “big data.” But there is not much business “synergy” in a self-driving car — it cannot drive synergetically. It has to drive for real, and not kill people and drive safely. There is no such a thing as adding business value to driving, the act and technology of driving itself. The greatest challenge for AI is driving safely, and there are some great research and development areas where humans are in the loop to assist. But generally, it’s both a Deep Learning and IoT problem -- AI is enabling a physical device which can drive and make life better in very tangible ways.
The third day is AI.vision where we will have industry leaders talk about the future of AI. We have one condition for our speakers: you can talk about the future if you can build the present. Most of our speakers are technical founders who grew through the ranks and now have executive leadership positions where they define strategy and build stuff themselves. They work with teams who build stuff, so they know what goes inside of AI in production.
In the possible future, where do you see the breakthroughs in AI coming, and how will that change people’s lives?
Everybody is trying to do bots now. I have seen some tremendous science which goes into natural language processing, so one of the areas in which we will see breakthroughs is language understanding. It’s super hard, so I don’t think that the majority of bot companies will cut it. I see tremendous science in NLP research where for decades people took the time to invest in language understanding. For example, Berkeley’s Semantic Machines is such a company. There will be a drastic difference between people who have decades of research under their belts versus those people who don’t, because language is capable of everything, and you need a whole set of approaches to model the way we can model with language. It will be a big step for AI to have bots that can maintain conversation on a wide spectrum of topics, so-called “general intelligence,” and I think it is coming quickly.
Richard Socher, the Chief Scientist of Salesforce, and a speaker that day, spoke at the previous Data By the Bay conference on how all AI problems can be reduced to Q&A. We also have Stuart Russell as keynote on Human-Compatible AI, the area where we need to focus to make sure AI obeys our values in a provable way. Jeremy Howard, the author of a well-known Deep Learning course, will also speak about computer vision for developers, another way to enable more of us to build systems understanding the world. It’s a really exciting way to learn to build AI.
