This past Thursday night, the Cognitive Artificial Intelligence Meetup hosted a talk by Dr. Dave Ferrucci, the leader of the team behind IBM’s Watson and the founder of a new company, Elemental Cognition, which is seeking to develop the next iteration of Artificial Intelligence [AI].
Taking place at the Ebay offices on 6th Avenue in New York City, the event was packed with probably 150–200 attendees, many of whom are seasoned AI practitioners — data scientists, programmers, engineers, architects and other pro’s much smarter than me.
What I expected to be a more casual interview turned out to be a brisk and dense lecture, one where I was struggling to keep up, furiously writing down words I didn’t understand and blown away by the ease with which they flowed out of this brilliant scientist and were absorbed by an equally brilliant audience.
AI lies at the intersection of many different disciplines — computer science, data science, statistics, learning theory, neuroscience, game theory, even philosophy — so without the same background or working knowledge, it can be extremely intimidating. Ultimately, though, the lecture was a pretty thorough primer on the biggest themes and challenges in the field, and I’d like to attempt to translate and distill some of what I learned to give you all a peak into this fascinating world.
We all know that computers are excellent at drawing conclusions from huge volumes of data, but the real insight comes from examining how they go about drawing these conclusions. It’s important to remember that AI is still beholden to human programmers, so the quality of its results depends on the data we allow it to access and how we tell it to find the answer.
The professor described some different ways we can use programming to find answers using a pretty simple question. Let’s start with Pattern Recognition.
Computers are Wrong
To review, one of the ways AI goes about predicting/answering questions is by combing through historical data, searching for conditions that look like the present day. If there’s a situation matching what we we’re experiencing — a similar combination of predetermined factors — then it’s likely that the outcome of that situation in the past will tell us what will happen in the future.
In the example Dr. Ferrucci highlighted in his presentation, the AI didn’t generate the most illuminating response, which isn’t so bad when we’re talking about rain boots. But what could be the consequences if the question is more serious than rain boots? Ferrucci used a striking example from his personal life to help paint a picture.
Some years ago, Dave’s father suffered a severe cardiac arrest. His father’s situation deteriorated to the point where the doctors feared he was brain dead and suggested that Dave sign a DNR — an order to not to revive a dying patient — effectively giving up on his Dad.
Dave refused, demanding to know why they believed they were right. The doctors, like AI, pointed to several statistical factors — his fathers age, the time between the incident and the beginning of treatment, etc. — but that wasn’t good enough. Dave asked for evidence, and the doctors cited dilated pupils; Dave wanted to know what other factors — like his medication — might cause this condition. On and on this went, Ferucci picking apart the diagnosis, going through every step of his father’s care to find alternative solutions and requesting any test or treatment he could ask for. A few days later, his father was sitting up in bed, alert, and most definitely not brain dead.
This is not only an amazing story, but also a potent metaphor for the limits of AI. Pattern matching doesn’t provide a “causal model,” or put plainly, the system will only be able to tell us what it sees, not why it sees it. Just because things seem similar, does not mean they are the same. Just because two things happen at the same time, it doesn’t mean that one is caused by the other. And unlike the doctor, a computer isn’t able to answer deeper, more specific follow-up questions, only the ones we program it to answer.
And therein lies the rub. Do humans and computers really understand each other? How does this effect the way we use and treat AI? And how do we push the system beyond simple mimicry and regurgitation, and into a more collaborative, conversational model?
Watson, Language and the Beginning of Understanding
As mentioned before, Ferrucci led the team that created Watson, an AI that was able to whup our best champions in Jeopardy just a few years ago. He went into great detail about how they analyzed the game itself, but the real key to Watson’s genius lies in it’s ability to execute Natural Language Processing. After all, how can you find the correct answer if you don’t understand the question?
Since the clues in Jeopardy are varied and random, Watson couldn’t rely on a static approach and had go through a multi-step process to come up with the right answer. It first analyzed each statement, identified keywords, discovered how they related to each other (i.e. what was being asked), searched a database to find possible answers, and weighed those against each other to find the most likely correct response. For more complex questions, Watson broke them up them into smaller bits, ran through that process for each section, then stitched those answers together to form a composite response.
These breakthroughs not only led to an explosion in language-based applications, but also to Watson Paths, a research collaboration with the Cleveland Clinic that developed a set of technologies to assist medical professionals. Unlike with Jeopardy, medical diagnoses deal with much longer, more nuanced and complex questions. Instead of requiring someone to manually program a series of prompts — like in the first “Rain Boots” example — the system is able to find it’s own “path” toward a solution through a combination of analysis and machine learning. Paths asks and answers multiple questions, finds supporting evidence, and shows the steps it took to arrive at its conclusion using interactive visualizations and charts.
With Paths, we’re no longer just getting back a pattern, but actually seeing the critical reasoning process at work. The system was trained by doctors to mimic their process, and has learned it so well, that it is now being used to teach med students. At least in this one specific area, we seem to be taking real steps towards a collaborative understanding between human and machine.
Computers as Children
You can begin to see that AI is being built to mimic the same processes we use to figure things out for ourselves. An image Dave came back to over and over again was of a child beginning to take in the world around them.
We all know how annoying kids can be during this process, whenever they get new information, they ask “Why?” and keep asking until either the subject or the parent is exhausted. This is known as “building a shared model,” constructing a logical framework to understand a given concept. The complexity and power of computers is impressive, but when it comes to understanding natural speech with all of the inferences, subtleties and devices we take for granted, they’re toddlers. So why not treat them that way?
In explaining his newest venture, Elemental Cognition, and how it is going about developing the next stage of AI, Ferrucci leaned heavily on this concept in explaining how it’s teaching its technology a dreaded subject for anyone who’s taken standardized tests: Reading Comprehension.
The problem is several layers deep and the learning curve is a steep one. Ferrucci himself admitted that a true application of this technology is still a long way into the future, but the possibilities are profound — it would be analogous to moving from Siri to the computers on Star Trek. Instead of simply asking for small bits of information or the completion of simple tasks, we’d have systems that could adapt to not only the complexity of the questions we ask, but also to the ways in which we ask them.
Where Watson Paths is restricted to Medicine, it’s easy to imagine applications of learning/teaching/collaboration machines in everything from Law to Government, Engineering, Economics and probably a few fields that haven’t even emerged yet.
Takeaways with a Neuroscientist
As mentioned earlier, I couldn’t hang with most of the crowd after the lecture was over, but I managed to strike up a conversation with a neuroscientist who’s in the process of transitioning into Tech. One thing we discussed were the limits of language.
As complex and lush as it can be, there are many other ways we understand each other — intuition, emotion, things we simply can’t put into words — and by concentrating solely on language, AI is inherently missing a piece of the puzzle. And like humans, these AI could develop — if not a personality — a perspective that influences their decisions and actions. How could we compensate for these? Perhaps by developing a series of AI along a spectrum of biases and using each perspective in decision making.
And finally, the one I’ve been thinking about most, if we’re to train and develop these machines to mimic and help us, how will that process change the way we see ourselves? Will it reveal errors, advantages or perspectives we’ve never considered, and what will we do with that new knowledge?
We’ve covered a good bit of ground here and I’d love to know what you think about all these ideas and concepts, so let’s keep this going in the comments section below. And if you’ve got first hand knowledge or experience in AI, please feel free to let me know how I did with all of this and add your own insights to the conversation.
Be sure to keep an eye on the Tech2025 event calendar for more opportunities to wrap your mind around emerging technology, and if you want to get dig even deeper into AI, we’ll be hosting a workshop April 25th with Kathryn Hume of Fast Forward Labs: Explain It Like I’m Five: What’s the Difference Between AI, Machine Learning, NLP, and Deep Learning?
Until then. stay curious and stay informed.