Monthly Archives: August 2017

There is no shortage of A.I. researches leveraging the unique environments and simulations provided by video games to teach machines how to do everything and anything. This makes sense from an intuitive sense until it doesnt. Case in point: a team of researchers from Google DeepMind and Carnegie Mellon University using first-person shooters like Doom to teach A.I. programs language skills.

Huh?

Yes it sounds bizarre, but it works! Right now, a lot of devices tasked with understanding human language in order to execute certain commands and actions can only work with rudimentary instructions, or simple statements. Understanding conversations and complex monologues and dialogues is an entirely different process rife with its own set of big challenges. Its not something you can just code for and solve.

In a new research paper to be presented at the annual meeting of the Associate for Computational Linguistics in Vancouver this week, the CMU and DeepMind team detail how to use first-person shooters to teach A.I. the principles behind more complex linguistic forms and structures.

Normally, video games are used by researchers to teach A.I. problems solving skills using the competitive nature of games. In order to succeed, a program has to figure out a strategy to achieve a certain goal, and they must develop an ability to problem solve to get there. The more the algorithm plays, the more the understand which strategies work and which do not.

Thats what makes the idea of teaching language skills to A.I. using a game like Doom so weird the point of the game has very little to do with language. A player is tasked with running around and shooting baddies until theyre all dead.

For Devendra Chaplot, a masters student at CMU who will present the paper in Vancouver, a 3D shooter is much more than that. Having previously worked extensively at training A.I. using Doom, Chaplot has a really good grasp at what kind of advantages a game like this provides.

Rather than training an A.I. agent to rack up as many points as possible, Chaplot and his colleagues decided to use the dense 3D environment to teach two A.I. programs how to associate words with certain objects in order to accomplish particular tasks. The programs were told things like go to the green pillar, and had to correctly navigate their way towards that object.

After millions of these kinds of tasks, the programs knew exactly how to parse through even the subtle differences in the words and syntax used in those commands. For example, the programs even know how to distinguish relations between objects through terms like larger and smaller, and reason their way to find objects they may have never seen before using key words.

DeepMind is incredibly focused around giving A.I. the ability to improvise and navigate through scenarios and problems that have never been observed in training, and to come up with various solutions that may never have been tested. To that extent, this new language-teaching strategy is an extension of that methodology.

The biggest disadvantage, however, comes with the fact that it took millions and millions of training runs for the A.I. to become skilled. That kind of time and energy certainly falls short of an ideal efficiency for teaching machines how to do something.

Still, the study is a good illustration of the need to start introducing 3D environments in A.I. training. If we want machines to think like humans, they need to immerse themselves in environments that humans live and breathe in every day.

More:

Why FPS Video Games are Crazy-Good at Teaching AI Language - Inverse

AP Photo/Matt Rourke, File

WASHINGTON Facebook artificial intelligence bots tasked with dividing items between them have been shut down after the bots started talking to each other in their own language.

But hold off on making comparisons to Terminator or The Matrix.

ForbesBooks Radio host and technology correspondent Gregg Stebben said that Facebook shut down the artificial intelligence program not because the company was afraid the bots were going to take over, but because the bots did not accomplish the task they were assigned to do negotiate.

The bots are not really robots in the physical sense, Stebben said, but chat bots little servers or digital chips doing the responding. The bots were just discussing how to divide some items between them, according to Gizmodo.

The language the program created comprised English words with a syntax that would not be familiar to humans, Stebben said.

Below is a sample of the conversation between the bots, called Bob and Alice:

Bob: i can i i everything else

Alice: Balls have zero to me to me to me to me to me to me to me to me to

Though there is a method to the bots language, FAIR scientist Mike Lewis told FastCo Designthat the researchers interest was having bots who could talk to people.

If were calling it AI, why are we surprised when it shows intelligence? Stebben said. Increasingly we are going to begin communicating with beings that are not humans at all.

So should there be fail-safes to prevent an apocalyptic future controlled by machines?

What we will find is, we will never achieve a state where we have absolute control of machines, Stebben said. They will continue to surprise us, we will have to do things to continue to control them, and I think there will always be a risk that they will do things that we didnt expect.

WTOPs Dimitri Sotis contributed to this report.

Like WTOP on Facebook and follow @WTOP on Twitter to engage in conversation about this article and others.

2017 WTOP. All Rights Reserved.

Go here to see the original:

No, Facebook did not shut down AI program for getting too smart - WTOP

Script kid-ai ... What the malware-writing bot doesn't look like

Feature The magic AI wand has been waved over language translation, and voice and image recognition, and now: computer security.

Antivirus makers want you to believe they are adding artificial intelligence to their products: software that has learned how to catch malware on a device. There are two potential problems with that. Either it's marketing hype and not really AI or it's true, in which case don't forget that such systems can still be hoodwinked.

It's relatively easy to trick machine-learning models especially in image recognition. Change a few pixels here and there, and an image of a bus can be warped so that the machine thinks its an ostrich. Now take that thought and extend it to so-called next-gen antivirus.

Enter Endgame, a cyber-security biz based in Virginia, USA, which you may recall popped up at DEF CON this year. It has effectively pitted two machine-learning systems against each other: one trained to detect malware in downloaded files, and the other is trained to customize malware so it slips past the aforementioned detector. The aim is to craft software that can manipulate malware into potentially undetectable samples, and then use those variants to improve machine-learning-based scanners, creating a constantly improving antivirus system.

The key thing is recognizing that software classifiers from image recognition to antivirus can suck, and that you have to do something about it.

Machine learning is not a one-stop shop solution for security, said Hyrum Anderson, principal data scientist and researcher at Endgame. He and his colleagues have teamed up with researchers from the University of Virginia to create this aforementioned cat and mouse game that breeds better and better malware and learns from it.

When I tell people what Im trying to do, it raises eyebrows, Anderson told TheRegister. People ask me, Youre trying to do what now? But let me explain.

A lot of data is required to train machine learning models. It took ImageNet which contains tens of millions of pictures split into thousands of categories to boost image recognition models to the performance possible today.

The goal of the antivirus game is to generate adversarial samples to harden future machine learning models against increasingly stealthy malware.

To understand how this works, imagine a software agent learning to play the game Breakout, Hyrum says. The classic arcade game is simple. An agent controls a paddle, moving it left or right to hit a ball bouncing back and forth from a brick wall. Every time the ball strikes a brick, it disappears and the agent scores a point. To win the game, the brick wall has to be cleared and the agent has to continuously bat the ball and prevent it from falling to the bottom of the screen.

Endgames malware game is somewhat similar, but instead of a ball the bot is dealing with malicious Windows executables. The aim of the game is to fudge the file, changing bytes here and there, in a way so that it hoodwinks an antivirus engine into thinking the harmful file is safe. The poisonous file slips through like the ball carving a path through the brick wall in Breakout and the bot gets a point.

It does this by manipulating the contents, and changing the bytes in the malware, but the resulting data must still be executable and fulfill its purpose after it passes through the AV engine. In other words, the malware-generating agent can't output a corrupted executable that slips past the scanner but, due to deformities introduced in the binary to evade detection, it crashes or doesn't work properly when run.

The virus-cooking bot is rewarded for getting working malicious files past the antivirus engine, so over time it learns the best sequence of moves for changing a malicious files in a way that it still functions and yet tricks the AV engine into thinking the file is friendly.

Its a much more difficult challenge than tricking image recognition models. The file still has to be able to perform the same function and have the same format. Were trying to mimic what a real adversary could do if they didnt have the source code, says Hyrum.

Its a method of brute force. The agent and the AV engine are trained on 100,000 input malware seeds after training, 200 malware files are given to the agent to tamper with. These samples were then fed into the AV engine and about 16per cent of evil files dodged the scanner, we're told. That seems low, but imagine crafting a strain of spyware that is downloaded and run a million times: that turns into 160,000 potentially infected systems to your control. Not bad.

After the antivirus engine model was updated and retrained using those 200 computer-customized files, and it was given another fresh 200 samples churned from the virus-tweaking agent, the evasion rate dropped to half as the scanner got wise to the agent's tricks.

Sponsored: The Joy and Pain of Buying IT - Have Your Say

See the original post:

In the red corner: Malware-breeding AI. And in the blue corner: The AI trying to stop it - The Register