IBM AI researchers say ‘what is the question’ is the real question
IBM AI technology lead, John Smith, and principle research scientist Kush Varshney talk with ZDNet about how the company endeavors to broaden the discipline from “narrow AI” to something more fit for the rigours of industry. Defining what the actual problem is that one is trying to train neural networks for, seems to be a big part of what makes AI work or not work, they suggest.
There is a place somewhere between today's machine learning technology and some future "AI" that is murky and difficult and conflicted.
Into that breach, IBM endeavors to insert itself as a voice of competency and experience.
At the prestigious NeurIPS machine learning conference in Montreal this week, IBM executives John Smith, manager of AI technology at IBM, and Kush Varshney, a principal research scientist with IBM Research, were making the case that the company has a role in how a still very "brittle" machine learning field can be more reliable and "trustworthy," depending on what one means by that phrase.
"It's about moving from narrow AI, where all of this really powerful technology has been highly accurate, but within a limited area of application, and making it something broader, something less brittle and something explainable," Smith told ZDNet.
Perhaps not "Artificial General Intelligence," says Smith, but something that lies between that Holy Grail of AI and today's actual implementations of neural networks.
Some of that involves particular original technical achievements, which IBM researchers will discuss this week. For example, Lazaros C. Polymenakos, is presenting the paper Knowledge Grounded End-to-End Dialog, which builds on prior work in the field of sentence embeddings in natural language. It tries to give greater understanding to machine learning models by treating individual entities in natural language statements as having their own separate space in memory.
And then there's work that broadens out from the technical approaches to a notion of how one defines the problem in machine learning to begin with. An example is Project Debater, a computer system that engages people in a back-and-forth dialogue.
"We got onto this point of how can a computer puts together arguments," says a Smith. "Debater is about how does the computer come up with these problems in the first place, how does the computer go and do all of its homework?"
"The interesting thing about Debater is, it's not just reading but listening," says Smith. "Listening to how are people saying things, it's listening comprehension."
"If you are looking for a single, end-to-end model, it's not that, it's the whole NLP pipeline, and making all of that work in this situation, and establishing a baseline."
To Smith, again and again the matter of AI comes back to defining a problem. "You can come up with a great idea" from a technology standpoint, "but then have no data, or not have anything to do with the data that's interesting."
IBM has selectively picked areas where it believes it can help, such as cases of bias, for example. At IBM's booth at the show, the company displayed monitors describing efforts such as a an exploration of predictions of recidivism among felons in the US.
The research firm ProPublica had done a study, in 2016, of predictions of recidivism that was performed using a private algorithm, "COMPAS," sold to law enforcement by the firm called Northpointe. ProPublica's study had found that African Americans, for example, were predicted by the algorithm to have higher rates of recidivism than other members of the population, which was not actually supported by historical recidivism data -- an instance of race bias.
Varshney gestured to results on the display that showed IBM could revise such data to come up with predictions that were highly accurate, but also didn't have the high bias. Why this is hard, and why it benefits from IBM's technology, he says. "The difficulty is the dependency between the statistical elements" says Varshney. "You have to go through and find all the ways that there are dependencies that operate that cause bias even if you've explicitly removed these variables of race, gender, etc., or explicitly compensated for them." (More on the effort in the original blog post on the matter.)
Beyond such isolated incidences, Varshney says an in increasing issue he and others will tackle is the overall question of what makes statistics reliable. Machine learning is one application of statistics, and whether it is used narrowly or broadly, there is still the fundamental question of how well statistics can be relied on to make inferences about populations and behavior and future occurrences. Another example of putting things into practice is work conducted with Memorial Sloan Kettering Cancer Cancer on the "International Skin Imaging Collaboration melanoma project. (That work is separate from IBM Watson work being done with Memorial Sloan Kettering Cancer Center.)
Although the work is purely academic at this point, Varshney is optimistic, saying that "hopefully we will get it into the clinic soon." The idea of an AI-assisted doctor is a decades-old dream, he is aware, but "first we need to get the accuracy that's required - it comes back to trusting AI."
Such examples of partnership comes back to Smith's larger point of defining the problem.
"It's a conversation we have all the time with subject matter experts, when we talk to one another that's when ideas emerge."
He notes that with IBM having over 400,000 employees "in every industry ... we get a lot of real world problems coming back to us."