Artificial Neural Networks Help Hack Biological Counterparts

May 12, 2019 by Roger Cheng 16 Comments

Eyes are windows into the soul, the old saying goes. They are also pathways into the mind, as much of our brain is involved in processing visual input. This dedication to vision is partly why much of AI research is likewise focused on machine vision. But do artificial neural networks (ANN) actually work like the gray matter that inspired them? A recently published research paper (DOI: 10.1126/science.aav9436) builds a convincing argument for “yes”.

Neural nets were named because their organization was inspired by biological neurons in the brain. But as we learned more and more about how biological neurons worked, we also discovered artificial neurons aren’t very faithful digital copies of the original. This cast doubt whether machine vision neural nets actually function like their natural inspiration, or if they worked in an entirely different way.

This experiment took a trained machine vision network and analyzed its internals. Armed with this knowledge, images were created and tailored for the purpose of triggering high activity in specific neurons. These responses were far stronger than what occurs when processing normal visual input. These tailored images were then shown to three macaque monkeys fitted with electrodes monitoring their neuron activity, which picked up similarly strong neural responses atypical of normal vision.

Manipulating neural activity beyond their normal operating range via tailored imagery is the Hollywood portrayal of mind control, but we’re not at risk of input injection attacks on our brains. This data point gives machine learning researchers confidence their work still has relevance to biological source material, and neuroscientists are excited about the possibility of exploring brain functions without invasive surgical implants. Artificial neural networks could end up help us better understand what happens inside our brain, bringing the process full circle.

[via Science News]

Neural Networks Using Doom Level Creator Like It’s 1993

May 16, 2018 by Drew Risinger 12 Comments

Readers of a certain vintage will remember the glee of building your own levels for DOOM. There was something magical about carefully crafting a level and then dialing up your friends for a death match session on the new map. Now computers scientists are getting in on that fun in a new way. Researchers from Politecnico di Milano are using artificial intelligence to create new levels for the classic DOOM shooter (PDF whitepaper).

While procedural level generation has been around for decades, recent advances in machine learning to generate game content (usually levels) are different because they don’t use a human-defined algorithm. Instead, they generate new content by using existing, human-generated levels as a model. In effect they learn from what great game designers have already done and apply those lesson to new level generation. The screenshot shown above is an example of an AI generated level and the gameplay can be seen in the video below.

The idea of an AI generating levels is simple in concept but difficult in execution. The researchers used Generative Adversarial Networks (GANs) to analyze existing DOOM maps and then generate new maps similar to the originals. GANs are a type of neural network which learns from training data and then generates similar data. They considered two types of GANs when generating new levels: one that just used the appearance of the training maps, and another that used both the appearance and metrics such as the number of rooms, perimeter length, etc. If you’d like a better understanding of GANs, [Steven Dufresne] covered it in his guide to the evolving world of neural networks.

While both networks used in this project produce good levels, the one that included other metrics resulted in higher quality levels. However, while the AI-generated levels appeared similar at a high level to human-generated levels, many of the little details that humans tend to include were omitted. This is partially due to a lack of good metrics to describe levels and AI-generated data.

Example DOOM maps generated by AI. Each row is one map, and each image is one aspect of the map (floor, height, things, and walls, from left to right)

We can only guess that these researcher’s next step is to use similar techniques to create an entire game (levels, characters, and music) via AI. After all, how hard can it be?? Joking aside, we would love to see you take this concept and run with it. We’re dying to play through some gnarly levels whipped up by the AI from Hackaday readers!

Continue reading “Neural Networks Using Doom Level Creator Like It’s 1993” →

One-Pixel Attack Fools Neural Networks

April 15, 2018 by Elliot Williams 34 Comments

Deep Neural Networks can be pretty good at identifying images — almost as good as they are at attracting Silicon Valley venture capital. But they can also be fairly brittle, and a slew of research projects over the last few years have been working on making the networks’ image classification less likely to be deliberately fooled.

One particular line of attack involves adding particularly-crafted noise to an image that flips some bits in the deep dark heart of the network, and makes it see something else where no human would notice the difference. We got tipped with a YouTube video of a one-pixel attack, embedded below, where changing a single pixel in the image would fool the network. Take that robot overlords!

Or not so fast. Reading the fine-print in the cited paper paints a significantly less gloomy picture for Deep Neural Nets. First, the images in question were 32 pixels by 32 pixels to begin with, so each pixel matters, especially after it’s run through a convolution step with a few-pixel window. The networks they attacked weren’t the sharpest tools in the shed either, with somewhere around a 68% classification success rate. What this means is that the network was unsure to begin with for many of the test images — making it flip from its marginally best (correct) first choice to a second choice shouldn’t be all that hard.

This isn’t to say that this line of research, adversarial training of the networks, is bogus. The idea that making neural nets robust to small changes is important. You don’t want turtles to be misclassified as guns, for instance, or Hackaday’s own Steven Dufresne misclassified as a tobacconist. And you certainly don’t want speech recognition software to be fooled by carefully crafted background noise. But if a claim of “astonishing results” on YouTube seems too good to be true, well, maybe it is.

Thanks [kamathin] for the tip!

Continue reading “One-Pixel Attack Fools Neural Networks” →

Catastrophic Forgetting: Learning’s Effect On Machine Minds

June 23, 2017 by Steven Dufresne 22 Comments

What if every time you learned something new, you forgot a little of what you knew before? That sort of overwriting doesn’t happen in the human brain, but it does in artificial neural networks. It’s appropriately called catastrophic forgetting. So why are neural networks so successful despite this? How does this affect the future of things like self-driving cars? Just what limit does this put on what neural networks will be able to do, and what’s being done about it?

Numerical weights in an artificial neural network

Neurons in the brain

The way a neural network stores knowledge is by setting the values of weights (the lines in between the neurons in the diagram). That’s what those lines literally are, just numbers assigned to pairs of neurons. They’re analogous to the axons in our brain, the long tendrils that reach out from one neuron to the dendrites of another neuron, where they meet at microscopic gaps called synapses. The value of the weight between two artificial neurons is roughly like the number of axons between biological neurons in the brain.

To understand the problem, and the solutions below, you need to know a little more detail.

Continue reading “Catastrophic Forgetting: Learning’s Effect On Machine Minds” →

Creepy Speaking Neural Networks

March 22, 2017 by Elliot Williams 37 Comments

Tech artist [Alexander Reben] has shared some work in progress with us. It’s a neural network trained on various famous peoples’ speech (YouTube, embedded below). [Alexander]’s artistic goal is to capture the “soul” of a person’s voice, in much the same way as death masks of centuries past. Of course, listening to [Alexander]’s Rob Boss is no substitute for actually watching an old Bob Ross tape — indeed it never even manages to say “happy little trees” — but it is certainly recognizable as the man himself, and now we can generate an infinite amount of his patter.

Behind the scenes, he’s using WaveNet to train the networks. Basically, the algorithm splits up an audio stream into chunks and tries to predict the next chunk based on the previous state. Some pre-editing of the training audio data was necessary — removing the laughter and applause from the Colbert track for instance — but it was basically just plugged right in.

The network seems to over-emphasize sibilants; we’ve never heard Barack Obama hiss quite like that in real life. Feeding noise into machines that are set up as pattern-recognizers tends to push them to the limits. But in keeping with the name of this series of projects, the “unreasonable humanity of algorithms”, it does pretty well.

He’s also done the same thing with multiple speakers (also YouTube), in this case 110 people with different genders and accents. The variation across people leads to a smoother, more human sound, but it’s also not clearly anyone in particular. It’s meant to be continuously running out of a speaker inside a sculpture’s mouth. We’re a bit creeped out, in a good way.

We’ve covered some of [Alexander]’s work before, from the wince-inducing “Robot Bites Man” to the intellectual-conceptual “All Prior Art“. Keep it coming, [Alexander]!

Continue reading “Creepy Speaking Neural Networks” →

Train Your Robot To Walk With A Neural Network

December 11, 2016 by Elliot Williams 22 Comments

[Basti] was playing around with Artificial Neural Networks (ANNs), and decided that a lot of the “hello world” type programs just weren’t zingy enough to instill his love for the networks in others. So he juiced it up a little bit by applying a reasonably simple ANN to teach a four-legged robot to walk (in German, translated here).

While we think it’s awesome that postal systems the world over have been machine sorting mail based on similar algorithms for years now, watching a squirming quartet of servos come to forward-moving consensus is more viscerally inspiring. Job well done! Check out the video embedded below.

Continue reading “Train Your Robot To Walk With A Neural Network” →