New Artificial Intelligence Cracks CAPTCHAs
April Flowers for redOrbit.com – Your Universe Online
There have been movies and books around for quite a while depicting Artificial Intelligence (AI) software and machines that become self aware. Perhaps the most famous was The Terminator movie franchise, but it is by no means the only example. Until today, there seemed to be little hope of that ever happening. Vicarious, a California startup company that develops AI software, has announced that the algorithms used by its software can now reliably solve modern CAPTCHAs, including Google’s reCAPTCHA. Google’s reCAPTCHA is the world’s most widely used test of a machine’s ability to act human.
CAPTCHAs (Completely Automated Public Turing test to tell Computers and Humans Apart) are those little boxes with random phrases or words or numbers at the bottom of a web page’s login. Until now,the ability to use them correctly proved you were a human logging in, not a computer set to spam the company, because they are specifically designed to be easy for humans and hard for computer programs.
“Modern CAPTCHAs provide a representative snapshot of many of the problems encountered in generic visual perception: large variation among instances of objects, segmentation that requires understanding of the objects, and contextual disambiguation,” said Vicarious researchers.
“Our strategy is to solve CAPTCHAs using algorithms that are instances of a general framework for solving problems found in human perception and reasoning. This allows us to transfer our learning from solving CAPTCHAs to more general problems, like vision.”
If, according to a Stanford University study, an algorithm is able to reach a precision of at least one percent, a CAPTCHA scheme is considered broken by the machine. Vicarious’ AI uses core insights from machine learning and neuroscience to achieve a success rate of up to 90 percent on modern CAPTCHAs from Google, Yahoo, PayPal, Captcha.com and others, rending text-based CAPTCHAs ineffective as a Turing Test.
A Turing Test is an assessment of a machine’s ability to exhibit intelligent behavior equal to, or indistinguishable from, that of a human. Alan Turing introduced the idea in a paper titled “Computing Machinery and Intelligence” published in 1950. The original purpose of the Turing test was to determine whether a computer is able to fool a human interrogator into believing that it is also human, however, a newer “standard interpretation” is used today. The standard interpretation asks whether a computer is able to imitate a human. That is where Vicarious’ AI comes in.
“Recent AI systems like IBM’s Watson and deep neural networks rely on brute force: connecting massive computing power to massive datasets. This is the first time this distinctively human act of perception has been achieved, and it uses relatively minuscule amounts of data and computing power. The Vicarious algorithms achieve a level of effectiveness and efficiency much closer to actual human brains”, said Vicarious co-founder D. Scott Phoenix.
The Vicarious RCN is not the first to break CAPTCHAs, but it is the first to do it reliably and with such a high success rate. Early efforts to create CAPTCHAs were poorly designed and easy to solve, using Optical Character Recognition (OCR) or standard machine learning methods. Modern CAPTCHAs, like Google’s reCAPTCHA, are designed so well that these approaches have a zero percent accuracy. Single CAPTCHAs have been broken on a rare basis by researchers exploiting bugs or idiosyncrasies in the generation process; however, such simple programs don’t attempt to understand the image. The errors they exploit are easily patched by fixing the CAPTCHA generator.
“Understanding how brain creates intelligence is the ultimate scientific challenge. Vicarious has a long term strategy for developing human level artificial intelligence, and it starts with building a brain-like vision system. Modern CAPTCHAs provide a snapshot of the challenges of visual perception, and solving those in a general way required us to understand how the brain does it”, said Vicarious co-founder Dr. Dileep George.
Modern CAPTCHAs present a representative snapshot of many of the problems found in generic visual perception. For example, a large variation among instances of objects, segmentation that requires understanding of the objects, and contextual disambiguation. Vicarious’ research team set out to solve CAPTCHAs using algorithms that are instances of a general framework for solving problems found in human perception and reasoning, which allows them to apply the results of solving CAPTCHAs to more general problems, like vision.
According to Vicarious, their algorithm is general purpose and can be applied to many sensory perception and reasoning problems, not just CAPTCHAs. Previous studies have linked higher level reasoning with a grounded perception system. In other words, our mental models of the physical world are linked to our perception and action system. In addition, one-third of the brain’s cortex is dedicated to vision. Vision is easier to test and debug than other intelligences, according to the researchers, and provides immediate and obvious benefits across multiple fields when solved.
DOES THE AI WORK LIKE A HUMAN BRAIN?
Vicarious set seven criteria for judging whether an AI system works like a brain:
1. Solve problems that are easy for humans and very difficult for computers – Vicarious’ emphasis is on building systems that can both demonstrate human-like intelligence and map back to the cortex.
2. Require very few training examples to learn and generalize from new concepts – Human brains are very good at generalizing from a few examples, which is considered a hallmark of intelligence. Most modern learning algorithms require tens of thousands of examples, while the new Vicarious AI only uses between one and five training examples per letter.
3. Build its understanding of the world using sensory data, like a human does – The AI experiences the world as a human, through still and moving sequences of images, to build an understanding of common and uncommon shapes. It relates patterns it has experienced in the past to what it experiences in the present.
4. Exhibit roughly equivalent capabilities to a human with the same experience – Like a human brain, the Vicarious AI recognizes text-based CAPTCHAs at greater than 90 percent accuracy.
5. Have a detailed mapping to neurobiology and neuroanatomy – Whether or not an AI uses neuron-like computational elements, it should be possible to derive a detailed mapping of the algorithm to biologically plausible circuits. a biological mapping of the algorithms of the Vicarious AI is possible.
6. Explain multiple neurobiological phenomena – Once the mapping is complete, it should be possible to use that neurobiological map to draw conclusions about the purpose of unexplained phenomena to build confidence in the model’s ability to work like a human brain. The Vicarious AI explains several visual illusions and the role of feedback connections in the brain.
7. Generalize across multiple domains – In the human brain, the same network of connections is responsible for understanding all of our senses, as well as language, motor actions, and higher level conceptual reasoning. Special purpose computer programs cannot be easily repurposed to handle other tasks, such as handling motor actions, recognizing photographs, or reasoning about high-level ideas. A brain-like system should be flexible enough to transfer to other sensory modalities and problem domains easily.
COMPARED TO OTHER LARGE-SCALE BRAIN PROJECTS
Vicarious’ AI is one of several large-scale brain projects that are on-going at this time. Others, like the EU’s $1 billion Human Brain Project (HBP) or DARPA’s $100 million SyNAPSE project, are focusing on simulating the brain as much as possible. Projects like these are confronted with a mix of brain processes relevant to intelligence, depending on the level of detail. Such processes only exist because of biological necessities, like metabolism.
The researchers hypothesize that the information processes of the brain are too interwoven with those parts of the brain essential for its own communication and biological survival, so there is no single clean abstraction level upon which to work. They focused their research efforts on teasing out the essential elements for information processing, looking initially to the structure and organization of the world outside the brain.
IN THE FUTURE
The researchers say that solving CAPTCHA is the first public demonstration of Vicarious’ Recursive Cortical Network (RCN). The commercial applications of RCN, although still many years away, will have broad implications in the areas of robotics, medical image analysis, image and video search, and many other fields. They also say that their algorithm is general purpose and can be applied to many sensory perception and reasoning problems, not just CAPTCHAs.
“We should be careful not to underestimate the significance of Vicarious crossing this milestone,” said Facebook co-founder and board member Dustin Moskovitz. “This is an exciting time for artificial intelligence research, and they are at the forefront of building the first truly intelligent machines.”