Tuesday, January 10, 2017

Tech Tuesday: On the Cutting Edge of A.I. Research with Kathryn Hume of Fast Forward Labs

Our human fascination with the future is nothing new. 19th century dreamer Jules Verne's imagination ran far ahead of what 19th century science could pull off. Was that a nuclear-powered submarine Captain Nemo was commandeering? Isaac Asimov, sci fi and hard sci author of more than 300 books wrote about travel in outer space as well as the evolution of intelligent machines. Today we have whole career fields in areas that once lived only in the realm of imagination.

With the 2017 Consumer Electronics Show behind us, it seems a good time to introduce Kathryn Hume, Director of Sales for Fast Forward Labs, a leader in the realm of artificial intelligence.

EN: I like your slogan “Reporting on the Recently Possible.” It implies that you’re on the cutting edge. What are a couple examples of the “recently possible” that you have been sharing or involved with.

KH: By "recently possible," we refer to technologies for which it's more possible to build working, real-world software products today than it was one year ago, and for which it will be even more possible to build great products in the next few years. Our expertise lies in artificial intelligence (A.I.), so we focus our research on software that uses data (as opposed to other new technologies like the blockchain). And it's very exciting to work in A.I. these days, as it's recently gotten the attention not only of geeks interested in cognition, but also of the investment and enterprise community.

One trend that excites us is deep learning, a particular machine learning technique that does a great job processing rich, complex data like images and text. We've built two prototypes using deep learning. The first automatically classifies the pictures on a user's Instagram feed according to their content, enabling users to organize visual information in a different way than the chronological stream. This is similar to tools Google is offering to organize your photos according to who's in them (separating, say, friends and family). These techniques have wide applications for medicine (automated radiology), insurance (automated claims adjustments), self-driving cars (processing environmental data), and robotics (improving usability at home or in manufacturing lines). We also studied using deep learning to automatically summarize text, in our Brief prototype. This uses a slightly different deep learning architecture that can track long-term dependencies in a sequence of data. If we think about it, text is basically a sequence of letters combined together to make meaning: neural networks make mathematical models of language, and use properties of these models as proxies for the linguistic meaning we use to communicate with one another.

EN: What are some ways that businesses are applying A.I. today?

KH: The first generation of A.I. products largely consisted of recommender systems (with companies like Amazon recommending products based on past behavior of similar buyers) and classification systems (things like spam filters or sentiment analysis to help brands monitor consumer preferences). These past few years have opened up a host of new and exciting A.I. applications: we now see companies using things like natural language generation to automatically write factual news articles from structured data sets (like company earnings reports) or using conversational bots to answer standard customer service inquiries.

That said, A.I. is vastly overhyped. We find that most of the large enterprises we work with are just beginning to build solid data programs, and having a lot of data, formatted in a way that's digestible by algorithms, is a condition to succeed with A.I. Many companies are starting to form intuitions about processes they may automate using A.I., but simply lack the data required to train a system that will perform reliably. We encourage companies to think hard about collecting as much data as possible now so they can eventually mimic human processes using machines in the future. Once this happens, it could lead to automation of many routine processes, including those that seem to require specialized human judgment, like finding relevant documents for a lawsuit.

EN: What is currently happening in publishing in the realm of A.I.?

KH: One of the big publishing trends is automated article writing using natural language generation. Software can now write company earnings reports, local weather reports, sports reports, basic crime reports, and other factual, descriptive news. But these systems are far from being able to do interpretative journalism, and companies like the Associated Press have developed ethical standards for when it's right - and not right - to use natural language generation, given the nuances of including subjective viewpoints.

We are also seeing publishers explore automated summarization tools to provide readers as many articles as possible about a relevant topic. For example, we're working on a project with a financial news provider that will summarize news about a particular stock ticker for investors.

The biggest concern right now in publishing and A.I. are filter bubbles and fake news. Many algorithms based on recommender techniques optimize for user engagement, and end up showing people content that aligns with polarizing ideologies, curtailing our ability to be empathetic, democratic citizens. In the wake of the 2016 election, there are efforts across the A.I. community to develop tools that can identify false content and encourage higher-quality journalism in the new networked age.

EN: What are some ways that A.I. will make the world a better place?

KH: A.I. research is a wonderful, meaningful activity. The very process of developing systems that seem to think, that seem to be intelligent, forces us to ask deep questions about how our minds work, who we are, and what society we want to create and inhabit. I entered into this discipline because I was fascinated by the fact that computers processed meaning in language so much different than we do. Neural networks, the linear algebra matrices behind deep learning, transform words into long strings of numbers (vectors), adding them and multiplying them in ways that mirror analogies and predicate logic in words. How amazing is that? How amazing is it to see the abstract structure we can tease out of the messy world of communication?

And yet, we should not think that A.I. systems, just because they are computational, are somehow more objective and rational than we subjective humans are. A.I. systems are built on data, and data contains the traces of human actions and human society. There is a lot of attention being paid these days to the ethical pitfalls of A.I. systems, which can perpetuate social biases, denying credit to individuals based on socioeconomic factors, advertising lower-paying jobs to women than men based on historic salary data, or even recommending a longer sentence to a criminal based on their race. These algorithms are not evil. They are neutral magnifying glasses of the inequalities in society we don't want to admit exist. As with many challenging experiences in life, we need to have the courage to look these issues head on and do what we can to address them. Only then can we empower ourselves to use A.I. to create the society we want, not the one we have.

Besides that, A.I. will lead to wonderful conveniences and opportunities for discovery in our personal lives. We can use data to navigate cities we visit, allowing the system to find the restaurant we'd want based on our historic preferences or draw our attention to a movie we'd regret missing. Mobilizing trends across populations will improve cancer research, which currently suffers from a dearth of clinical trial information. Personalized medicine is another huge avenue that will take off in the next 10 years. Self-driving cars are poised to change how we live in cities. The possibilities are endless.


EN: What is the mission of Fast Forward Labs?

KH: Fast Forward Labs is an applied research company focused on commercializing A.I. research coming out of academia. We build prototypes that transform ideas presented in abstract academic research papers into tangible, real-world products, and write reports that explain how these technologies work. We also advise large enterprises and startups building cutting-edge products on their A.I. strategies, helping them apply new A.I. techniques in their own business environments.

EN: What will be the big A.I. story from the Consumer Electronics Show in Vegas this year? (Asked last week going into CES.)

KH: The big A.I. story at CES this year will exist somewhere in the realm of robotics, but likely be hidden from view. There will be an amazing product with an intuitive user interface that solves a big problem, and this product will be powered by data without the end user's even knowing it involves A.I. A good example of this is Google Maps, a seemingly boring, deceptively simple tool that has literally changed the way we navigate physical space. The backend includes complex data and engineering work we don't see, as we're focused on the tool solving the problem it was built to solve.

* * * *
Fast Forward Labs is based in Brooklyn, New York. Learn more and see the near future here.

These are very exciting times we live in. Let's see what happens next.

No comments: