What Is Pattern Computer?by Pilar Ackerman |
Pilar Ackerman, webmistress for Pattern Computer®, caught up with Chief Technology Officer Ty Carlson to get answers to the questions we’re being asked by journalists, potential customers, and investors.
OK, let’s start with the obvious question: What’s a pattern computer?
“Pattern Computer” is the name we have chosen to describe our company and our computing system, which is optimized across hardware and software for the discovery of patterns within complex, high-dimensional data.
What’s the significance of such capabilities, in layman terms? Why should anyone care?
The ability to discover new patterns can lead to valuable breakthroughs in science, health, and industry. Let me give some examples of some things we’re currently working on or have in queue, and readers can determine for themselves if they care.
Our work in breast cancer and lung cancer led us to the hypothesis that specific combinations of genes are at work simultaneously, when current insights typically consider a single gene by itself or at most a 2-way combination. And some of the gene associations we’re seeing include genes for which FDA-approved drugs exist — but for a different disease. That potentially means that a drug could be repurposed, and because it has already passed clinical trials for safety, it could be brought to market for this new use in an expedited manner.
We’ve just begun a partnership with a major cancer research hospital to supplement their methods with ours.
We are about to embark on a project looking at a combination of public and private data in the transportation industry to improve safety and efficiency.
The domain where the data originates is irrelevant to our system — medical, manufacturing, environmental — it doesn’t really matter. Now that’s a big statement, but it’s one I’m confident we’ll prove over time. But for now, we’re focusing on biomedical projects and we’re just starting to test our system outside that arena with the transportation project.
How is what you’re doing different from supercomputing? quantum computing?
Supercomputing is really a hardware-based arms race. Procure or build the fastest processors, connect a gazillion of them with a fabric, employ extreme parallel programming techniques, source a ton of faster memory, and find some place on the power grid capable of supporting your energy needs. I know this is an oversimplification, but that’s really the core of it. Common “brute force” techniques on a supercomputer are not enough to do what we do.
Quantum computers exploit the special nature of quantum particles which can represent multiple states (any weighted combination of 1 and 0). In contrast, a classical computer can only represent a binary state (a 1 and a 0). This creates a huge, complex range of possible outcomes which makes possible the simultaneous computation of all possible solutions to a problem and is orders of magnitude faster than classical computers.
There’s a lot of interesting progress being made, but I think both the makers and the industry at large agree that it’s not yet ready for prime time. By the time it is, we plan to be ready and able to leverage it. But it’s not something we’re focused on today.
Let’s talk about our Pattern Computer system on the hardware side first. There’s nothing we’re currently doing on the hardware side component-wise that isn’t commercially available, but the cool thing is that our hardware design will allow us to add heterogenous processor types, all operating at the same time. One advantage is that as new HW is developed, we don’t risk our system becoming obsolete. Another is a performance boost, based on our ability to automatically route specific data to the best available processor(s) for the task at hand. I think that’s a pretty smart model.
It is in the software where the real magic happens. Some of our algorithms are proprietary, and some are modifications of existing techniques being applied in a new way. I won’t provide more detail on this, but the net result is a system specifically designed and optimized for pattern discovery. To date, I haven’t come across another company that has the same focus on overcoming the industry challenges surrounding the “discovery” of unknown interactions & insights in data, in quite the same way or on the same scale as Pattern Computer.
Building on this last point, how would you describe the difference between traditional pattern recognition and pattern discovery?
In the case of pattern recognition, you already know what you are looking for and you have created a template or model of some type that helps you find what you are seeking. Facial recognition is a great example of this, as is the “fingerprint” of a data security breach. It essentially does what a human can already do, but at very high speeds across huge data sets.
Pattern discovery differs in that it presents to the user the most important relationships, many of which will have never been seen or documented before. It’s possible these discoveries might have eventually been made via serendipity, but to my knowledge there’s never been a solution that systematically discovers patterns and presents them for consideration. We know this will drive new hypotheses, and it’s the way I believe that computers will add the greatest value in the future.
Does it actually work? If every discovery is new, how would you even know if it’s working correctly?
One way that we test it is very interesting — almost a high-tech version of hide-and-seek. One portion of our engineering team embeds patterns of their choosing in otherwise random data, and another portion of our team finds them the using the system. These are huge test sets with up to hundreds of thousands of rows and millions of columns, so it’s a non-trivial problem and a valid test.
The way we develop our software also gives us confidence. We always start with the math. Once we know the math works, we begin to code, first the prototype, then the highly optimized, scalable version. We then go through a variety of tests, including the hide-and-seek method I just mentioned.
You used the words, “every discovery is new” so I want to clarify. Every discovery we make is not necessarily new. What we find are the most important relationships. Some of them may have been discovered already by serendipity or years of research. I would be concerned if there were a complete absence of known results in our output — that would be a signal that something is probably wrong.
You asked if it works. Yes, it does. We believe the computational results are accurate (we can go back from the result set to the locations of the original data objects) and we’re in the process of validating our results biologically. Once we do that, we are off to the races.
Talk to me about that 7-minute video posted on the www.patterncomputer.com website — the one that looks like planets in the night sky.
That’s the result of some recent work that we’ve only disclosed to current and potential investors until now. We looked at mutated genes in lung cancer patients. Each dot is a mutated gene that exists in 1% or more of the population of patients studied. The red dots are genes for which an approved drug treatment for lung cancer exists. Not a lot of red dots are visible, which is a tough realization if you have lung cancer. And there are no patterns apparent in the raw data, at least not to human eyes.
When the video cuts to what looks like constellations, with the planets linked by lines — now that’s where it gets really interesting. Our pattern discovery system can find and display patterns in the data and display the topology of the gene associations. In some cases, they are pairwise, in other cases three-way associations, and sometimes they are much higher order. If we color the genes for which a drug exists but is approved by the FDA, but for a disease other than lung cancer (we used yellow), we get new insights that could inform approaches like repurposing the drug for lung cancer. Or perhaps combining it with the drug that already exists for lung cancer for the red gene would yield good results. The statements I’m making are all hypotheses, not yet proven biologically. Our system is really a hypothesis generator, providing new ideas to investigate and explore. Does it guarantee that each pattern it finds will be super-meaningful in its given field of study? No, of course not. But what could we do with a machine that identifies important, promising things to study? Now that’s a great question, and that’s why we’re here. Stay tuned!