Startup creates Artificial Intelligence that can read Captchas and in 5-7 years plans to put human intelligence into mathematical algorithms

The world’s first artificial intelligence (AI) to pass the Completely Automated Public Turing test to tell Computers and Humans Apart — Captcha — debuts today from Vicarious FPC Inc. in San Francisco.

Vicarious is a three-year old California Flexible Purpose Corporation (thus the FPC after its name), which instead of maximizing shareholder value, like a normal corporation, is aiming to fulfill a singular purpose — to solve the important algorithmic problems behind building a human-like AI. For Captcha, its already succeeded, but Vicarious’s long-term goal is to generalize its Captcha AI into a complete robotic brain that is as smart as a human in all areas of sensory perception.

“Our goal is to combine insights from neuroscience with modern machine learning techniques and cast them into mathematical algorithms that are just as intelligent as humans,” Phoenix told us.

The three-year-old startup was founded by Phoenix, formerly entrepreneur in residence at Founders Fund, and co-founder Dileep George, formerly chief technology officer of Numenta. Vicarious is running on its second round of funding. The first seed round in 2010 was $1.1 million, and its Series A for $15 million was just completed last year, giving the six-person company the time and money it thinks it needs to fulfill its mission of casting human intelligence into mathematical algorithms, which it expects to achieve in five to seven years.

Vicarious – Turing Test 1: Captcha from Vicarious Inc on Vimeo.
Text based captchas – strengths and weaknesses (14 pages, 2011)

Vicarious’s secret weapon is what it calls a recursive cortical network (RCN) — a machine learning framework that embodies the structural and computational power of the brain’s neocortex.

“Our algorithms express what we think is going on inside the neocortex of the brain,” says Phoenix.

The big difference between RCN and conventional neural networks, is that it does not try to mimic the biological mechanisms inside the brain, but instead mimics the hierarchical representations of information in the natural world, which it claims are mirrored in the functionality of the brain.

“The brain has evolved to mimic the structure found in the physical world,” said George, whose brainchild is RCN, in an interview with EE Times. “Our algorithms are modeling a direct correspondence between the structure in the neocortex of the brain and the statistics of real-world data.”

George claims that the structures found in the brain not only mimic the statistical regularities found in nature, but that they are “recursive” — meaning that the same basic structures are reproduced identically at all hierarchical levels in the brain, from individual neurons and synapses all the way up to complete brain centers, such as the vision and speech centers.

Thus the big difference between Vicarious’s models of the brain and the neural network models of the colossal efforts of the European Union’s Blue Brain Project and the US Defense Advance Research Project Agency’s (DARPA’s) Systems of Neuromorphic Adaptive Plastic Scalable Electronics (SyNAPSE), is that Vicarious is not trying to mimic all the details of the biological brain, but only its overall functionality.

Vicarious claims its algorithm is actually more complex than traditional neural networks, but not in the details of physical brain structures. Instead, they omit what they call the superfluous details of how the brain is interconnected, and instead concentrate on duplicating the repeating scaffold at each level of brain function.

“We have a very detailed mapping of what our algorithm is doing in relation to the various layers of the cortex — it’s a one-to-one mapping to each layer of the brain — and in that way is more detailed that any other algorithm of its kind,” says George. “But we are not modeling all the little details, such as how the brain uses spikes to pass information from neuron to neuron.”

Vicarious claims that its algorithm models learning in individual neurons in the same way as it does high-level concepts by making assumptions about how the world works, such as that “water makes things wet” and “you need an umbrella to go out in the rain,” says Phoenix. “Understanding language, for instance, uses the same circuitry as vision uses to perceive objects, and that will eventually allow our algorithms to truly understand the world and the language humans use to describe it.”

Vicarious claims it already has the evidence it needs to prove its RCN is truly modeling the brain’s functionality, namely, that it can generalize from just a few examples while learning, rather than requiring many, many examples before providing reliable results, as is the case with conventional artificial neural networks.

“Our algorithm learns the same way a child does, allowing it to recognize new instances of an object after only about 10 examples, whereas traditional neural networks might need 1,000 or even 10,000 examples before they can generalize well,” says George.

Beyond Captcha

As Vicarious progresses down its path of mimicking human intelligence, it plans to release more Turing Test milestones, first in the area of general visual perception, such as picking out specific objects from video footage, then branching out into the other senses, such as natural language recognition, and eventually to a complete robotic brain that encapsulates a human level of general intelligence.

“Breaking Captcha was just a stepping stone along the way to solving general human intelligence,” says Phoenix. “Along the way there are a lot of things we hope to achieve that computers can not do today — from identifying cancers in x-rays to automating manufacturing to a robotic brain for Rosie-the-Robot from The Jetsons — that is our eventual goal.”

If you liked this article, please give it a quick review on ycombinator or StumbleUpon. Thanks