Last week we reached an exciting milestone in Compellon’s history with the closing of our recent round of funding. While we are a startup company delivering a new platform for “insights as a service,” the technology underpinnings of our platform are the result of my three-decades long research and development as a practicing “data scientist” (yes, long before that term was invented)! As one of the founders of the company and its Chief Scientist, I’d like to share a bit of the Compellon back-story, and kick off what I hope will be an ongoing blog on how Compellon today is delivering toward an exciting vision of the future of artificial intelligence.
Long before founding Compellon and prior to immigrating to the US, I built advanced statistical analysis systems for companies both as a professor and a researcher in the Soviet Academy of Sciences. The most intriguing part of my work was the investigation of projects which all pursued reasonable objectives, had sufficient funding, were well equipped, and were conducted by a team of highly qualified specialists…and in spite of all this—failed. I saw such examples across many fields: manufacturing, environmental monitoring, medical diagnostics, acoustical testing of materials, and others. Figuratively speaking, my job in such projects was “autopsy” and “resurrection”. Overall, the main culprit in the overwhelming majority of cases was the use of inadequate assumptions based on well-established beliefs.
Nowadays, many companies call their approaches assumption-free, and I feel like the term has achieved near buzzword status as a concept that is obviously good. But this direction, in fact, was not so obvious in the mid-70s and early 80s. Theories for the goals and evolution of artificial intelligence were mushrooming into many varied and innovative directions. Those that resonated with me addressed head-on the challenges I’d experienced dealing with faulty assumptions and how to avoid them. One proposed direction focused on balancing man-machine labor division through theories for automatic generation of hypotheses based on evidence. Another from industrial automation called for autonomous decision devices. And engineering psychology made a major quest for intelligent control entities able to work under conditions where humans can’t.
I was active in these circles and heavily influenced by the style of multi-disciplinary collaboration popular at that time among Eastern European scientists. I absorbed an exciting mix of ideas from a wide range of areas: probability theory and statistics, information theory, approximation theory, abstract algebra, algorithm theory, topology, mathematical logic, different parts of physics, several engineering disciplines, and psychology. It inevitably led me to the concept of autonomous intelligent agents (framed in current terminology). To quickly summarize, the objectives were approximately the same as articulated by Daniel Curtis (see definition of “Artificial Psychology” in Wikipedia): 1) agents make autonomous decisions based on new, abstract, and incomplete information, 2) they can reprogram themselves based on data, 3) they can resolve programming conflicts originated from incomplete or contradictory data, and 4) all of this is not incorporated in advance in the original code or operating program. These were—and still are—the keys to creating intelligence autonomously from the data and removing the need for human assumption.
My goal was to combine my interest in the general understanding of AI evolution with the practical analytical needs of engineering, environmental, social, and other everyday activities. (This contrasted another popular AI motif to create entities which exhibit capabilities of humans, such as man-machine chess competitions and so forth.) I realized that to take the challenge seriously, without resorting to half-measures and lightweight solutions, the focus should be on creation of practically useful agents making autonomous data-driven decisions that are capable of self-awareness (which would lead to self-improvement ability). In other words, it would be machine learning based on the strong AI agenda, where the I in AI would stand for intelligence complementary to humans.
This AI research direction, while exciting and innovative, unfortunately stalled at the time due to political upheaval that interrupted research funding in the former USSR, along with many tough, unsolved technical challenges and the limits of the compute power of the day.
Meanwhile, worldwide artificial intelligence and its subfield, lately dubbed as “machine learning,” have evolved along an alternative path based on extreme fragmentation of tasks, algorithm optimization, and brute force from ever growing computing power. It is exciting how solutions have matured by leaps and bounds and are used increasingly in mainstream applications. But my ideal goals of autonomous, data-driven decisions have not yet been achieved, as the state of the art is still heavily reliant on human effort, expertise – and assumptions. (Though interestingly enough, from reading recent academic papers I see that the old themes mentioned earlier once again are coming back into vogue in research circles.)
I continued working on a cross-discipline inspired methodology for self-aware agents after immigrating to the US (as always, in parallel with their practical application). Along this path, I was fortunate to meet Ken Charhut and Mark Pollard, my Compellon co-founders. One of the inspirations for our collaboration was the application of my methods to a post-mortem clinical study of a promising new device Ken’s company was developing, which demonstrated only a lukewarm effectiveness. After my analysis, it turned out that the reason for such inconclusive results was reliance on conventional wisdom in the cardiac discipline which was just plain wrong in the face of the evidence in the data. If the trial had been structured differently, the device would have demonstrated much better outcomes. Ken and Mark immediately saw the promise of how a truly assumption-free, 100% evidence-based approach could transform pretty much any line of business. Together we formed Compellon to incorporate the science I’d been applying in one-off consulting projects into a modern platform architecture that can be applied to any data, by any user.
We’ve now spent several years building these capabilities into our technology platform, Compellon 20|20. An exciting side effect of autonomous data-driven decision capability is the dramatic simplification of the user experience versus any other system on the market, precisely because our approach only requires users to have business questions and data to explore, but not to be an analytics expert. This is what most excites a great deal of our users.
And that is of course, in my view, exactly the way it should be.
Now, for those who intersect in the technology side of how we are able to do this, I look forward to sharing more about our approach and vision for the evolution of machine learning based on strong AI in my future blogs. And if you haven’t engaged Compellon yet, but are interested in how a true assumption free approach is transformative in helping you solve practical problems, please contact us to check out the Compellon20|20 platform with your data.