I just got back to my hotel room after attending the first of a two day Cognitive Computing Forum, a conference running in parallel to the Semantic Technology (SemTech) Business Conference and the NoSQL Conference here in San Jose. Although the forum attracts less attendees and has only a single track, I cannot remember attending a symposium where so many stimulating ideas and projects were presented.
What is cognitive computing? It refers to computational systems that are modeled on the human brain – either literally by emulating brain structure or figuratively through using reasoning and semantic associations to analyze data. Research into cognitive computing has become increasingly important as organizations and individuals attempt to make sense of the massive amount of data that is now commonplace.
The first forum speaker was Chris Welty, who was an instrumental part of IBM’s Watson project (the computer that beat the top human contestants on the gameshow Jeopardy). Chris gave a great overview of how cognitive computing changes the traditional software development paradigm. Specifically, he argued that rather than focus on perfection, it is ok to be wrong as long as you succeed often enough to be useful (he pointed to search engine results as a good illustration of this principle). Development should focus on incremental improvement – using clearly defined metrics to measure whether new features have real benefit. Another important point he made was that there is no one best solution – rather, often the most productive strategy is to apply several different analytical approaches to the same problem, and then use a machine learning algorithm to mediate between (possibly) conflicting results.
There were also several interesting – although admittedly esoteric – talks by Dave Sullivan of Ersatz Labs (@_DaveSullivan) on deep learning, Subutai Ahmad of Numenta on cortical computing (which attempts to emulate the architecture of the neocortex) and Paul Hofmann (@Paul_Hofmann) of Saffron Technology on associative memory and cognitive distance. Kristian Hammond (@KJ_Hammond) of Narrative Science described technology that can take structured data and use natural language generation (NLG) to automatically create textual narratives, which he argued are often much better than data visualizations and dashboards in promoting understanding and comprehension.
However, the highlight of this first day was the talk entitled ‘Expressive Machines’ by Mark Sagar from the Laboratory for Animate Technologies. After showing some examples of facial tracking CGI from the movies ‘King Kong’ and ‘Avatar’, Mark described a framework modeled on human physiology that emulates human emotion and learning. I’ve got to say that even though I have a solid appreciation and understanding for the underlying science and technology, Mark’s BabyX – who is now really more a virtual toddler than an infant – blew me away. It was amazing to see Mark elicit various emotions from BabyX. Check out this video about BabyX from TEDxAukland 2013.
At the end of the day, the presentations helped crystallize some important lines of thought in my own carbon-based ‘computer’.
First, it is no surprise that human computer interactions are moving towards more natural user interfaces (NUIs), where a combination of artificial intelligence, fueled by semantics and machine learning and coupled with more natural ways of interacting with devices, result in more intuitive experiences.
Second, while the back end analysis is extremely important, what is particularly interesting to me is the human part of the human computer interaction. Specifically, while we often focus on how humans manipulate computers, an equally interesting question is how computers can be used to ‘manipulate’ humans in order to enhance our comprehension of information by leveraging how our brains are wired. After all, we do not view the world objectively, but through a lens that is the result of idiosyncrasies from our cultural and evolutionary history – a fact exploited by the advertising industry.
For example, our brains are prone to anthropomorphism, and will recognize faces even when faces aren’t there. Furthermore, we find symmetrical faces more attractive than unsymmetrical faces. We are also attracted to infantile features – a fact put to good use by Walt Disney animators who made Mickey Mouse appear more infant-like over the years to increase his popularity (as documented by paleontologist Stephen Jay Gould). In fact, we exhibit a plethora of cognitive biases (ever experience the Baader Meinhof phenomenon?), including the “uncanny valley”, which describes a rapid drop off in comfort level as computer agents become almost – but not quite perfectly – human-looking. And as Mark Sagar’s work demonstrates, emotional, non-verbal cues are extremely important (The most impressive part of Sagar’s demo was not the A.I. – afer all, there is a reason why BabyX is a baby and not an fully conversant adult – but rather the emotional response it elicited in the audience).
The challenge in designing intelligent experiences is to build systems that are informative and predictive but not presumptuous, tending towards the helpful personal assistant rather than the creepy stalker. Getting it right will depend as much on understanding human psychology as it will on implementing the latest machine learning algorithms.