AI Based on the Mammalian Visual Cortex can Easily Solve CAPTCHAs

Over the years, computer algorithms have grown increasingly better at recognising patterns, which has allowed them to sucessfully identify such objects as animals and human faces.

Impressive as that is, algorithms are still incapable of matching images to semantic meaning, and frequently run into problems when images they’re supposed to detect are distorted or obscured by surrounding noise. This is why many websites are still keen on using CAPTCHAs.

This, however, might not be a viable option for long – a California-based start-up called Vicarious had just announced its new algorithm, inspired by the mammalian visual cortex, which can easily crack trusty old CAPTCHAs with very little training.

The visual cortex arranges a scene in a hierarchical fashion by first identifying individual patterns, such as edges and surfaces (carried out by specialised groups of neurons), and then figuring out, based on proximity, which of those features belong to the same object.

CAPTCHAs might not hold out long into the future as effective means of determining whether or not the user is a human. Image credit: Jake Spurlock via, CC BY 2.0.

This way of organising visual perception allows mammals to recognise known objects even if they’re obscured or in a different orientation.

In mimicking the brain’s approach to pattern recognition, the company had developed a top-down algorithm, called the Recursive Cortical Network (RCN). It operates by first identifying contours, then surface features (such or smoothness) based on those contours, and finally arranging the recognised properties into pools based on proximity.

Once that is done, the individual pools begin to communicate with each other to calibrate feature choices, eventually resulting in groups of related features.

The algorithm then forms “object hypotheses” (collections of features that might be separate entities) and assigns them a specific score. After a number of rounds of assessing the highest ranked hypotheses against other contenders, the RCN can identify objects even if they’re moderately distorted.

Once deployed to solve CAPTCHAs, the algorithm performed slightly better than humans (94 percent letter recognition accuracy versus 87 percent). It’s enough for software to solve CAPTCHAs at one percent accuracy for the system to become useless from a security standpoint.

The RCN also did fairly well with the BotDetect, PayPal, and Yahoo systems, demonstrating 57 percent accuracy.


Comment this news or article