When you think Cognitive Computing, it’s likely that IBM Watson is one of the first things that come to mind. Certainly there’s good reason for that to be the case, as the vendor has done a remarkable job getting the market familiar with the term and the possibilities. Watson’s debut on Jeopardy remains a popular as well as computing culture highpoint.
But there’s a wider world of Cognitive Computing out there, represented by more than one vendor. Among these are companies such as DataRPM and its Cognitive Data Discovery technology; Google’s recently acquired artificial intelligence startup Deepmind; and, Saffron Technology, which was founded in 1999 and in March raised another $7 million in funding to continue to advance its business and its Associative Memory technology.
There may be a generally accepted idea of what Cognitive Computing is – computers that learn, think, and interact with each other and with other humans, underpinned by technology such as natural language processing, machine learning, and neural networks. But the nascent space is still home to more than one interpretation of how to realize these concepts to analyze information, answer questions, and drive decisions in a data-soaked world.
Indeed, recently-named Saffron Technology Chief Product Officer Ian Hersey likens Cognitive Computing to Big Data in the sense that both are labels that encompass many different technologies: “Big Data has gotten to the point where everyone’s got Big Data and Big Data strategies, but if you ask five different people what that is you’ll get five different answers,” says Hersey, whose background includes co-founding SAP-acquired text analytics vendor Inxight Software, and serving as Chief Product Officer at InTTENSITY and CTO of Attensity. “It’s probably a similar case with Cognitive Computing and what exactly makes systems cognitive.”
In Saffron’s case, Hersey explains, the answer lies with its Associative Memory approach to organizing data, which is similar to how the human brain approaches the information organization task in order to draw conclusions and take some action. That is, like humans, Saffron’s technology turns information into analogies to support a reasoning layer that is brain-like in its ability to understand patterns and make decisions based on that understanding. It auto-detects the similarity of new things to things previously seen or spots the anomalies, and then measures these similarities or dissimilarities at the query/reasoning level.
“If something is like something we saw before, we know what to do with it,” says Hersey of how humans process data – or if it’s different, it sticks out to us as something to do something about. “You can express that as rules but humans don’t put things through a rules engine.” Machine learning, on the other hand, can help with jobs like identifying outliers from normal behavior in a non-random way.
In addition to that, Cognitive Computing for Saffron also involves the ability to be as adaptive as humans, to automatically respond to new data that changes the system’s understanding:
“Humans don’t have derived statistical or semantic models in our heads,” he says. “We just get new data and new experiences in and that automatically becomes part of the information we use every day to make decisions.”
This is all in contrast to IBM Watson, he says, which, as a question-answering system that sits atop a base of user-entered facts, requires an intensive knowledge-engineering approach to tune it to support different domains like healthcare. Saffron instead deals with unstructured and structured data “that doesn’t exist in what you think of as a fact base,” reveals Hersey. It can store attributes and associations between data naturally, he says, whether that data is being ingested from smart devices, databases, or unstructured notes that will be parsed into structured data through its natural language processing component.
Ingested entities are indexed into Saffron’s Memory Base, around the associations they have with other entities. As an example, an incoming wave of press releases from Saffron might mention product names, corporate locations, key business partners, employees and such, which all would be indexed as having an association with Saffron.
“So a memory is created for Saffron and all the things it is connected to and the context around that,” Hersey says. “But each of those other elements is indexed as their own memories as well, so if I am mentioned in one release I get a memory that associates me with Saffron and with other things not necessarily related to Saffron.”
Saffron technically calls this a hypermatrix because it is a very large set of sparse matrices. In a sense this is equivalent to a graph database, Ian noted, but matrices are a more efficient way to represent graphs and perform computations on them. One of the things Saffron’s founders figured out is how to compress a large set of sparse matrices, which makes the computation of the connections and the strength of the associations between those connections “very efficient when we actually query to do similarity and anomaly detection,” Hersey says. “So the way we organize it is a big part of the secret sauce in how we return similar and dissimilar things very quickly.”
Hersey also explains that the strength of memory associations between two entities isn’t always parallel. For example, Saffron’s association to Hersey himself might be very strong, but his to Saffron’s might be weaker because he may have many more associations beyond his link to his employer. Think of it like this: A Twitter user may closely follow hundreds of celebrities, but have only 50 followers himself. In that case, his association to the celebrities would be stronger than any of those celebrities’ association to him, because these stars likely have more followers than people they are following.
What is the business value of this take on Cognitive Computing? Similarity finding is a big one, he says, giving aircraft maintenance as an example: “A big headache for airlines is unscheduled maintenance because when a plane is not flying the airline is not making money,” says Hersey. Being able to address an issue quickly could help avoid that, but how?
You can start things off by leveraging the ability to deal with the natural variations in language in text notes like past problem and solution descriptions – which may be described in different ways depending on the airline, airport location, or even whether a mechanic was in a rush or not – and identifying the structure within that data, to parse out part numbers, places or other information that could be useful in proactive maintenance activities and add them to the Memory Base.
Now, if a pilot gets an alert while flying that a part on the plane needs to be replaced, the stage is set for that part to be waiting when the plane gets to the gate. While not every airport has in stock the complete inventory of aircraft parts for every model plane, a lot of planes share similar structures, so another part may be able to do the job. “How do they quickly find the most similar one that will work if they don’t have the specific part number that is needed?” Hersey says.
The complexity of planes means parts inventories are very large, and finding the most similar part against a massive inventory requires you to be able to very quickly and in real-time look at the attributes and associations they have in common, Hersey says – such as whether they were used to solve similar maintenance issues, have they been used on this type of plane before, and so on. “That is all context around if something is similar or not,” he says. You can’t build a model in advance for that, “but it flows naturally as a result of how we index everything.”
He also sees a growing need for this technology as smart devices proliferate and analyzing the data from them becomes a high priority for more organizations – much as it already is in the medical field. There, for example, Saffron has worked with Dr. Partho Sengupta, a leading cardiologist at Mt. Sinai Hospital in New York City. He needed to use its HealthCare Intelligence Platform to distinguish patterns emerging from echocardiograms to determine whether they revealed a normal condition or one of two heart conditions that require very different treatments.
“If you look at patterns coming out of the machine with the human eye, they look very similar,” Hersey says, but making the wrong call and prescribing the inappropriate treatment can be fatal to the patient. Saffron loaded the data of 15 patients each for normal heart conditions and for the two diseased conditions, with some 10,000 attributes per heartbeat per patient, into its Associative Memory database, and let its Cognitive Computing capabilities go to work to result in 90 percent accuracy in discerning which patients had which conditions. Most physicians manually assessing this data have a 65 percent accuracy rating.
Cognitive Computing in the style of IBM Watson, based on a analyzing a question, generating a hypothesis of what the answer needs to look like, and then trying to find the best match from its facts base, doesn’t provide the tools to distinguish between the two conditions, when you don’t know in advance which attributes or combination of attributes will be important in for determining one condition from another, Hersey says.
Building more apps around the technology that can demonstrate its power and make it easy to use to more sectors is critical, he believes: “We have to make it easy for them to get the data in and do something very important to them, whether that’s saving lives or money or reducing gate wait time for airlines,” he says. Essentially, “we need to democratize this technology.”