What is Cognitive Computing?
[The following is a video and excerpt from the transcript of Susan Feldman's keynote address at KMWorld 2015. Susan Feldman is long-time technology analyst and the CEO of Syntexis Cognitive Computing Consortium. To see the full transcript of the video, click here
. You can view the video of the keynote at the bottom of this page.]
What is Cognitive Computing
What goes into cognitive computing? There's a certain amount of natural language involved. That's absolutely true. It's probabilistic, it's non-deterministic and it drives a lot of IT people crazy because they expect the answer, not some possible answers. It's very iterative. It's conversational. It's contextual. It learns in-depths about not only to you but about everybody else using the system, and it keeps getting smarter.
Machine learning is a requirement. It usually has a big data knowledge base. That means that if you have lots of data, even though somebody is living in the long tail in terms of his needs or profiles, you still have enough data, enough evidence to be able to understand that individual. You're not just aiming your business at the center of the bell-shaped curve. Analytics are built in various kinds and these technologies are highly integrated. It's not services-oriented architecture, though that's also important. They influence each other.
What Do Cognitive Systems Do?
What do these systems do? They analyze big data. They understand human language on multiple levels. They uncover relationships across sources. They understand and filter by context. They find patterns in the data that you didn't know existed. They find the black swans among the white swans, the surprises.
This is extremely valuable to competitive organizations, governments, and individuals. They learn from new information and new interactions. As you use the system, it gets better--we hope.
Whatís the next leap? For example, we can find drugs in the database of drugs and even side effects of drugs that are useful for controlling diabetes, but what's the best drug in this circumstance for this patient? We're looking for best, not just a list.
Or: Who's funding this terrorist organization, and how are the funds delivered? This is also a really useful question and one we've been asking for a long time. How big a threat is this organization? Can you make a recommendation about the level at which you need to react to what's going on?
Another example: Can you identify the most risky product or customer problems? It's very difficult to know what to pay attention to in the world of information overload and big data.
That's the kind of thing that you would want a bunch of humans to sit down and discuss, but there's an awful lot of information to weigh in through. Can we make the system a partner? Can the system make some recommendations, and then sit down and discuss it? We'll probably find stuff that we didn't know.
What Cognitive Computing Isn’t
Now we know what cognitive computing is. But what is it not?
Cognitive computing is more than big data or artificial intelligence. There are books coming out that say that machine learning is cognitive. I disagree, and so did the group of experts we convened to define this topic. It's not robotics. It's not drones. It's not humanoid. It's not entirely autonomous.
At least, that's what we feel right now. It is not a singularity and it's not a replacement for humans. It's an aid. It's another tool for us so that we can begin to understand our world and our problems and solve them more effectively.
What Does a Cognitive System Look Like?
Your senses take in clues from all over and then they discuss what they have found with other people in innovation. Both of those are really important but, those synapses of those connections are something which is very hard to mimic.
A traditional information system is very linear and sequential. You ask the question, it goes into the index, you've indexed all the documents, the documents get matched to the question usually with the terms in the question, the system outputs the information, and then you make a decision. The problem is, there really isn't enough interaction. It's almost as if the designers felt that people couldn't be trusted and therefore we had to have a wall between what we had in our heads and what they have in the system.
In a cognitive system, instead of a direct question you have a problem statement or exploration, and there's a great deal of iteration. One of the things that was a huge breakthrough for Watson on Jeopardy was analyzing the question. Before they ever threw the question into the system, they defined what kind of question it was and what kind of answer might be required. They had hundreds of different question types for Jeopardy. This is what people do too, right? If I ask someone in Washington, D.C., "How do I get to Reagan?" You're not going to tell me stuff about President Reagan because you're a person and you know what my context is.
By analyzing who, what, where, when types of questions, by extracting the elements of those questions, we are already ahead. More than that, when we ask a query we usually focus it down and make it as specific as possible. We have to because we're going to get garbage otherwise. That's not what we really want to do. At this point, especially in innovation, we want to explore. If we're exploring, we want to expand and create hypotheses. Weíll use each of those hypotheses to collect evidence to support or to deny that particular hypothesis.
That's what these systems do--particularly the Watson-like ones. We send in an expanded query or problem statement to a system that has done the orchestration of how to answer that particular kind of question. Different kinds of questions are more effectively answered by different combinations of tools. You may have four categorizers in one of these systems. How that question is answered depends on the type of question, which array of categorizers is most useful, and which one needs to have priority.
After weíve done the question analysis and expansion, it goes into a cognitive processor which is somewhat analogous to the index that hits the information store and does the similarity matching. Out comes, perhaps, a huge pile of information. But confidence scores have been assigned, evidence has either proven or disproven some of the hypotheses, or perhaps they've been combined. Thereís lots of stuff going on in there. There are sometimes voting algorithms to help with the confidence scoring. For Jeopardy, there was game theory, which is why you got such crazy wagers like, I'll wager $727 on this question, which is not a very human kind of thing to do. But apparently, Watson knew that it could afford to lose $727 but not 728. Anyway, outcome the data set and the filters are the who, what, where, when of who the person is, where they are in the task, what they've asked before.
The hypotheses are filtered down and sent to the exploration loop. The exploration loop is a set of tools that help visualize and analyze the information. At each step, what the system has done in terms of interaction, question answering, filtering, and analyzed data sets gets thrown back into their cognitive processor, making it smarter, more dynamic. That's important too because it follows you along in the process.
Because computers don't have your senses, in order to accomplish this, we have an array of technologies that work together. These include facial recognition, rich media understanding, content-intelligent services, machine learning, real-time voice translation, speech recognition, and taxonomies. Any tool that's useful to solve a particular kind of problem needs to be bundled into these systems.
Right now, at the dawn of the cognitive age, we're still trying to figure out what configurations work best for which problems. That's the kind of research that I'm working on right now because people keep asking me, what kind of problem is this good for and what do I look at? What are the tools?