AI Startup Says It Has Defeated Captchas

Shameless plug http://www.boyter.org/decoding-captchas/ I wrote this a few years ago. Has full source code too.

Advances in storage and compute have led to a disturbing fetishization of machine learning.

While the modern Machine Learning Movement makes sense in a historical context and is a reasonable reaction to the disappointing returns from symbolic inference during the early days of AI research, it is terrifying that the research community is satisfied to rely on big data and statistical methods to carry us forward.

Few among us recognize the need to prioritize the study of the human brain. Even fewer are placing their bets on intelligent computer systems seeded with neurologically-inspired designs.

Vicarious gets it.

How long before others see the writing on the wall?

Now is the time to stop reacting. Now is the time to consider the field in a broad context and develop a balanced, holistic approach.

Consider this a wake-up call.

http://blog.perfmode.com/the-noml-movement/

It generally means you teach it to recognize 3d shapes (a 2d image that moves = a 3d image, more or less. Yes there's a good reason why you might want to call it 2.5d, but the easy way to model a 2.5d object is in 3d). Think of it as the difference between recognizing 2 points and recognizing a Feynman diagram.

This is one of the things people don't often realize you can do with algorithms. You don't need to look at the world the way it actually really exists, and there may be very good reasons not to. Training algorithms to actually recognize moving images is incredibly hard, because it requires things like memory, fade-outs, recurrent networks, all that very advanced stuff. Obviously time exists as a continuum in the "real" world. But that's bloody inconvenient. So just look at big "quanta" of time, collecting all data points during the quantum, analyse it, then shift the quanta/window ahead 0.1s and do the exercise again. This is so much easier you wouldn't believe it.

Teaching an algorithm to recognize, say, a car collision, given 100 frames. It doesn't require any change to the algorithm (just a change in training data). And obviously your backend system needs to be aware that, over time, the "isColliding" output will look like ......1.....11.....1111...1111.1.1.111.11..11...11.11...11...1.....1...1...... when a collision occurs and this of course doesn't mean you've had 20 collisions.

It does mean a bigger network, slower training, and more resources needed. But not as much as you'd think. Keep in mind that a "temporal" network will need more hidden layers. Also please consider building "redundant" networks for temporal data. When people ask why, I have no better answer than that it's the same technique our brain uses, so frankly if it's good enough for God, it's good enough for me.

Doing the temporal thing means you're back to using trivially simple algorithms, running on more data.

Cracking captcha's is not very impressive. I've done it as a weekend project, and exceeding "average" human captcha'ing ability is easy. I actually got it to the point where my algorithm was slightly better at captcha's than me, where I was allowed to take 2 minutes for difficult captchas. If I wasn't allowed to take more than 10 seconds, my algorithm easily beat me by over 10% (my captcha performance, when measured, shockingly is only ~83%). I didn't cheat : I used an external site's captchas (from dns.be).

The algorithm used was dead simple backpropagation.

> Eventually, the spammers will make a bot to analyze the distance between dots, group them into letters, and the race will be on to use other methods. I see this as a never ending virus/immunity battle. We're pretty much at the end of Captchas. Other methods like mouse movement, surfing time, scrolling, etc... can all be mimiced as well. Computers can or will be able to simulate humans very well, even our imperfections.

Love your optimism. I'd like to say that captchas are very limited : they have to be solvable by idiot humans. Captcha algorithms have no such limits they have to abide by. Since the anti-spam side of things is blocked at a certain point in the arms race, the other side is bound to win.

Why not just require, say, a google or facebook login and transfer the "eliminate spammers" problem onto them ?

Not really, unless your corpus consists mainly of hopelessly distorted characters.

They state a captcha solving rate of around 90%.

For OCR to be cost-competitive, you typically need it to be correct on about 98% of characters or more; below that and it is typically cheaper to have a human typing in the text than to have a human correct OCR'd text.

Modern OCR engines typically do better than 99% on text that isn't really badly damaged (my MSc. dissertation was on error correction in OCR, and as part of that I tested some engines with pages that had been crumpled, intentionally damaged with sand and liquids, and even then many of the engines managed more than 99%).