For many years it seemed as if the term Artificial Intelligence was a dirty word. It had the connotation of great promise with little commercial applicability. Today it seems that AI has not only found its cache’, it has been co-opted. The blockbuster games released all promise new AI and talk about it as a differentiator for their game experience. Everything from search systems to path planning algorithms is now trumpeting AI. Many of these groups are leveraging very old technologies and jumping on the buzz word bandwagon. Games, in particular, are still leveraging finite state machines and Nav Mesh capabilities (see my last blog). But there has been real progress.
So what is the buzz really about? It seems to me that there are really two big things happening in the industry and neither one solves the most challenging issues faced by Artificial Intelligence researchers. The first, is the solution promised by “big data” and enabled by increases in processor (CPU) power. Projects focused on “deep learning”, as recently discussed in MIT Technology Review (TR), are peeking interest in what AI might be able to do with all the data being collected and new powerful computers. Google and IBM have invested heavily in these areas. For Google the idea is to leverage the cloud with their hundreds of thousands of machines and Petabytes of data to get better recognition of content found in videos or images. How, for example, does the search system identify videos containing cats? It turns out that this kind of categorization is extremely difficult, particularly when there are thousands of categories. Broader categories (cats vs. Siamese cats), predictably, offer better results. Google’s major advance shows 16% recognition for larger category sets and 50% for reduced sets of categories (again this TR reported data). This is fascinating for several reasons: first, the amount of hardware and data put forward is vast, yet your average 6 year old can beat it hands down. Second, it demonstrates just how difficult the problem is, and why it is unlikely to be solved by brute force. The other dominant player here is IBM with its Watson system. The Watson system again applies massive amounts (although small compared to Google’s cloud) of hardware (2880 processors and 14 Terabytes of RAM) but it also takes a multi-strategy approach to the question answering (Q/A) it performs. What is interesting to me is that each strategy looks at the data from a different perspective. For example, one strategy for playing Jeopardy! was to find puns. So although there is some brute force computation occurring, there is also the application of individual strategic expertise. In this case, the strategy for recognizing a pun. Human beings with a great amount of skill programmed these capabilities into Watson.
The second major push for AI is the steady progress of statistical approaches to things such as Speech Recognition, Handwriting Recognition and general pattern recognition of faces and objects. The math is not for the lay person but is based on Hidden Markov Models or Neural Network approaches. These models have been refined over the last few decades through numerous additional techniques to improve their performance. These capabilities have enabled technologies such as Siri and other digital assistants to relate to humans in a more natural way. It also enables technologies such as the Microsoft Kinnect to recognize you and load your profile when you play a game. Some believe that these statistical technologies can be pushed further to solve more difficult problems. I, for one, am very skeptical of this approach. In fact, the success of Watson provides some evidence for my skepticism. Although Watson uses massive amounts of data a processing. It combines this with carefully constructed strategic expertise programmed by a team of talented and intelligent researchers. Many of these strategies are specific to the goal of winning at Jeopardy!
The complexity of intelligence, therefore, lies not in the data but in the myriad creative strategies that we as humans can apply to that data. Each strategy is specific knowledge in its own right and each must work in combination with others to enable intelligence.
Discovery Machine, Inc. has been focused on the problem of capturing and deploying specific expertise. Every industry, every plant, every lab, every training center, every office has experts that have learned strategies to solve problems. As much as industry would like to believe that these strategies can be boiled into generic statistical approaches, we know they cannot! Decades of attempts have failed to provide a single generic strategy for expertise. Discovery Machine provides an alternative. Capture your organizations expertise and only leverage the data that is needed to realize that expertise.