AI and Enterprise Knowledge Integration: Part 1
Artificial Intelligence may well be the most potentially transformative technology since the Cloud, but it’s clearly become the reigning champion for Tech hype and media buzz.
IBM’s Watson – a “cognitive” computer capable of answering natural language questions - was developed to compete on Jeopardy, a popular quiz show. In 2011, Watson competed against world champions Brad Rutter and Ken Jennings before a TV audience of millions…and beat them. At the end, Jennings remarked: "I for one welcome our new computer overlords".
In fact, the Watson that won Jeopardy was an outcome of decades of research in “Symbolic AI”. It used the knowledge representation and reasoning capabilities of Prolog, an AI language invented 45 years ago. IBM researchers painstakingly combined numerous proven language analysis algorithms with new approaches from statistical machine learning, backed with massively parallel hardware designed for the outstanding new software.
However excessive the ensuing hype, Watson’s victory did have the merit of blasting AI research out of its long winter hibernation and into AI 2.0 - the new science of statistical machine learning, especially Deep Neural Networks (DNN), whose algorithms are “trained” with huge data sets.
Progress came fast, as products like Alexa (Amazon’s conversational virtual assistant) hit the market, but expectations inflated even faster. In May 2016, Gartner fearlessly (and wrongly) predicted that by 2018 more than three million workers would be supervised by robot bosses. In August 2017, a sensational (but false) story surfaced that a panicked Facebook had shut down a pair of AI robots after they became dangerously smart and invented their own language!
Then, in the journal Nature of October 19, 2017, Google introduced AlphaGo Zero, a DNN for playing the complex Chinese game Go. Unlike its predecessor AlphaGo which was trained with data from human games, the new algorithm knew only the rules and learned by playing 5 million games over 3 days against itself. It then promptly dispatched several Go champions as well as its software predecessor.
While comments by the Google team were admirably measured, AlphaGo Zero provoked a new blast of media buzz about a self-learning “superhuman intelligence” that made “humans seem redundant”.
In 2018, with AI hype more intense than ever before, what should we make of all this excited speculation? And what are the implications for digital transformation?
AI 2.0: Time for a Reality Check!
The last few years have seen extraordinary progress in machine learning, especially DNN. Tech giants rely on these technologies to run their “born in the cloud” businesses, and traditional companies are putting them to work in digital transformation.
Nonetheless, AI 2.0 today still resembles (as one expert wryly observed) “spreadsheets on steroids”, with some important limitations… which its leading practitioners are the first to point out.
AI 2.0 is “narrow” not “general” intelligence
Oren Etzioni, CEO of the Allen Institute for AI, says that in popular imagination AI is “general” with “machines doing what humans do”, while in practice AI can handle only “narrow tasks once thought to require human intelligence”.
Algorithms are optimized for specific tasks and, as the AI Index notes, even a small change in the task can cause performance to plunge.
In circus parlance, AI is still “a one trick pony”. The pony does its trick really well… but then the show is over.
Deep Neural Networks are crude models of human learning
DNN is modeled on our understanding of how the brain learns, establishing synaptic connections between neurons. A fascinating and powerful approach, but some caution is in order:
o Neuroscientists warn that our understanding of learning is still rudimentary. DNN algorithms are only first level approximations. The human brain is multiple orders of magnitude more complex than any DNN.
o DNN works well only with homogeneous data, whilst enterprise information is heterogeneous, comprising documents, databases, images, voice, business rules… Integrating all this data into usable business knowledge is something humans can do (albeit at considerable cost in terms of time) but DNN cannot.
For the moment, biology still trumps electronics.
Deep Neural Networks can’t reason
AlphaGo Zero playing Go may look like superhuman reasoning, but it’s mostly just statistics and probabilities.
As Yann LeCunn, Chief AI Scientist of Facebook and leading DNN pioneer remarked, we still don’t have “learning machines that can reason not just perceive and classify”. DNN doesn’t have the internal “models of the world” that constitute true knowledge and enable human reasoning.
As LeCunn bluntly put it, the most sophisticated DNN “doesn’t have the commonsense of a rat.”
Now don’t get me wrong. AI research is moving fast and things will change. The important thing today, however, is that there are already an amazing number of use cases in different business sectors. The most urgent opportunity by far is making machine learning much easier to use in “smart applications”, with tools like the Codex AI Suite. A lot of business value is at stake, including many potential quick wins.
Even so, we should look ahead to other AI technologies businesses need for digital transformation, as they move beyond the automation of mostly routine, “narrow” tasks towards the augmentation and assistance of human intelligence in complex knowledge intensive work.
The challenge of Knowledge Integration
The celebrated management guru Peter Drucker coined the term “knowledge work” in 1950 and famously predicted that raising the productivity of knowledge workers would be the n°1 challenge for 21st century enterprises. Here are some examples of knowledge intensive work where AI technologies could be a big help:
- Having a 360° view of a business and its ecosystem before proposing the best financing solution
- Configuring a complex industrial product to meet precise customer requirements
- Analyzing a voluminous contract to identify key actions and risk factors for ongoing surveillance
- Searching through numerous judicial precedents to prepare a legal brief
- Organizing the outsourced maintenance of locomotives sold to train companies
- Ensuring compliance by a financial institution with multiple regulatory regimes
- And designing new products and services
Unlike the “narrow” tasks of AI 2.0, most knowledge intensive work involves heterogeneous data (often in different formats, scattered across the enterprise and its ecosystem) and requires reasoning. “Narrow”, by the way, is a description not to be confused with something unimpressive or lacking in value. For example, recognizing a cancer tumor from medical imagery is expert work that DNN does well, yet the data is homogeneous and the task is by nature extremely specific.
The biggest obstacle to using AI for knowledge intensive work in most big companies is their fragmented “knowledge landscape”. Data is everywhere but is highly heterogeneous (databases, documents, spreadsheets, business rules...), suffers from “semantic incoherency” (i.e. contradictory meanings) and lives in autonomous silos. Companies are a gold mine of information, but many find it nearly impossible to access and connect data from across different departments and regions.
This is a huge problem for leveraging AI in knowledge intensive tasks because, as Kirk Borne, Principal Data Scientist at Booz Allen Hamilton, says “Knowledge is about connecting the dots”.
To integrate this information, to “connect it up” as knowledge usable by computers, we need help in particular from the older tradition of “Symbolic AI”, where knowledge and reasoning take center stage.
Bear in mind that I’m not talking theoretical futures here. Combining AI 2.0 with Symbolic AI is something Tech giants like Google as well as some leading edge industrial and financial companies, are already doing.
On our side, we use a progressive and pragmatic approach - “Enterprise Knowledge Integration” - that I’ll explain more in my next post.