Boffins discuss AI space program at hush-hush IARPA confab
IBM, MIT, plenty of others invited to fill Uncle Sam's spy toolchest, but where's Google?
The US government is taking artificial intelligence research seriously again, and so are some companies that will surprise you.
IARPA – an agency whose job is to develop and furnish the US spy community with advanced technology – has gathered companies and academics to discuss modern machine intelligence at a one-day conference in Washington D.C.
This conference, held on Thursday, will steer the future of research into Machine Intelligence from Cortical Networks (MICrONS). Experts were asked to propose MICrONS-related projects that aim to "create a new generation of machine learning algorithms derived from high-fidelity representations of cortical microcircuits to achieve human-like performance on complex information processing tasks."
Attending companies also discussed some of the challenges of today's machine intelligence approaches and some of the techniques needed to surmount difficult artificial intelligence challenges.
The goals of the MICrONS program are so ambitious the scheme has more in common with a space program or some of IARPA-counterpart DARPA's "grand challenges" rather than a traditional research project.
According to one document, the goals are to:
- Propose an algorithmic framework for information processing that is consistent with existing neuroscience data, but that cannot be fully realized without additional specific knowledge about the data representations, computations, and network architectures employed by the brain;
- Collect and analyze high-resolution data on the structure and function of cortical microcircuits believed to embody the cortical computing primitives underlying key components of the proposed framework;
- Generate computational neural models of cortical microcircuits informed and constrained by this data and by the existing neuroscience literature to elucidate the nature of the cortical computing primitives; and
- Implement novel machine learning algorithms that use mathematical abstractions of the identified cortical computing primitives as their basis of operation.
In plain English, the program hopes to figure out a bit more about how the brain works on a biological level, with a particular emphasis placed on stuff like how neurons interact and how large sets of them are tied together via a brain superstructure commonly called "the connectome".
Participating researchers will also try to take what they learn from this research and implement it in software to create better algorithms for things like image analysis.
'No one is claiming we're going to solve how the brain works'
"Everybody there was battle-hardened in the field - you're not going to see any kind of wide-eyed bushy-tailed optimism, you have people who have been in this for a very long time," one attendee told The Register on condition of anonymity.
"No one at the conference is claiming we're going to solve how the brain works or image recognition as a class of problems".
One perplexing thing about the conference was the lack of much public participation from artificial intelligence hothouses Google, Facebook, and Microsoft. Instead the attendees were drawn from a mix of startups, universities, and IBM, which has a large-scale cognitive research effort.
Though the list of attendees isn't available, we were able to get hold of a list of the companies that gave presentations or informal chats at the conference.
These participants included companies such as: IBM, Qelzal Corp, Nvidia, Lambda Labs, Neuromorphic LLC, Numenta and Neurithmic Systems LLC.
And researchers from the following institutions were scheduled to turn up: Harvard and the Harvard Medical Center, SRI International, the Georgia Tech Research Institute, Rice, Rochester Institute of Technology, Downstate Medical Center, Oxford, Yale, Johns Hopkins, Washington University, Howard Hughes Medical Institute, Australia National University, Simons Foundation, University of Tennessee, University of California, George Mason University, Columbia, Arizona State University, University of Vienna, Baylor, Columbia, Princeton, UC Berkeley, UCLA, and MIT.
At the meeting, some of the things discussed were the types of challenges IARPA can evaluate companies on. One idea is for "some form of scene decomposition or scene clustering," a source told us. This involves picking out repeated objects in scenes and requires the AI system to be able to develop abstract representations of objects.
Another aspect is to develop a more theoretically rigorous understanding of what goes on in our brains when we think, and work out how to implement this digitally.
Some attendees we spoke to described the projects as doable but at the very limits of our understanding, whereas others were more skeptical of their practicalness. One thing is for certain – after years in the funding wilderness, the US government is again waking up to the possibilities of true general artificial intelligence – just don't mention the AI Winter.
"The intelligence community puts money into it, the government puts money into it," said one attendee. "The environment is pretty warm right now, it's a new spring." ®