#neural nets

[eldritch-elrics]

love wins

[vs-stardream]

here are my kirby dall-e adventures

[sparkylurkdragon]

DEATH ON VIOLENT WINGS

DEATH TO EVERYTHING

[eka-mark]

Delivering malware covertly and detection-evadingly is critical to advanced malware campaigns. In this paper, we present a method that delivers malware covertly and detection-evadingly through neural network models. Neural network models are poorly explainable and have a good generalization ability. By embedding malware into the neurons, malware can be delivered covertly with minor or even no impact on the performance of neural networks. Meanwhile, since the structure of the neural network models remains unchanged, they can pass the security scan of antivirus engines. Experiments show that 36.9MB of malware can be embedded into a 178MB-AlexNet model within 1% accuracy loss, and no suspicious are raised by antivirus engines in VirusTotal, which verifies the feasibility of this method. With the widespread application of artificial intelligence, utilizing neural networks becomes a forwarding trend of malware. We hope this work could provide a referenceable scenario for the defense on neural network-assisted attacks.    

[allthingslinguistic]

Exploring Massively Multilingual, Massive Neural Machine Translation

Posted by Ankur Bapna, Software Engineer and Orhan Firat, Research Scientist, Google Research “... perhaps the way [of translation] is t...

An interesting blog post from Google AI/Google Translate about machine translation drawing on multiple languages at the same time, which can help make better translations, especially for languages that have fewer resources to draw on. Excerpt: 

Over the last few years there has been enormous progress in the quality of machine translation (MT) systems, breaking language barriers around the world thanks to the developments in neural machine translation (NMT). The success of NMT however, owes largely to the great amounts of supervised training data. But what about languages where data is scarce, or even absent? Multilingual NMT, with the inductive bias that “the learning signal from one language should benefit the quality of translation to other languages”, is a potential remedy. Multilingual machine translation processes multiple languages using a single translation model. The success of multilingual training for data-scarce languages has been demonstrated for automatic speech recognition and text-to-speech systems, and by prior research on multilingual translation [1,2,3]. We previously studied the effect of scaling up the number of languages that can be learned in a single neural network, while controlling the amount of training data per language. But what happens once all constraints are removed? Can we train a single model using all of the available data, despite the huge differences across languages in data size, scripts, complexity and domains? In “Massively Multilingual Neural Machine Translation in the Wild: Findings and Challenges” and follow-up papers [4,5,6,7], we push the limits of research on multilingual NMT by training a single NMT model on 25+ billion sentence pairs, from 100+ languages to and from English, with 50+ billion parameters. The result is an approach for massively multilingual, massive neural machine translation (M4) that demonstrates large quality improvements on both low- and high-resource languages and can be easily adapted to individual domains/languages, while showing great efficacy on cross-lingual downstream transfer tasks.

Read the whole post. 

[intelligentliving]

An AI Can Draw What You Look At By Decoding Your Brainwaves

A pair of neural networks reconstruct human thoughts from their brainwaves in real-time to draw what they are seeing in realtime with just a headset.

[wolframalpha]

An Intro to Neural Nets

Neural Nets aren't the same as brains, just inspired by them. [image: woman's profile & network graph] Like how birds inspired planes, or burrs inspired Velcro.

A neuron in the human brain hears signals from thousands of other neurons, listens to each excitatory or inhibitory input, and synthesizes the great mass of information from all of them to yield... [image: neuron illustration]

... blip , or no blip. [image:  line graph with a spike on the y axis labeled "Fig. 1 'Blip.'" and a line graph with a slight bump on the y axes labeled "Fig. 2 'A lack of blip.'"]

Neural nets in machine learning are like that, just cleaner. [image: network graph with arrows]

After all, neurons in the human brain had to cobble together their structure over millennia of evolution, using neurotransmitters, ion channels, and precise voltage control to achieve... [image: illustration of an ion channel]

... basically what we can do with a plus sign on a computer. [image: woman pressing a key on her laptop at her desk *bloop*]

Because neural nets are experiencing so much attention right now, they're going through a bit of evolution of their own. [image: network graphs with walking feet climbing onto land] New ways of constructing neural nets are coming out every day, and bringing with them...

... lots and lots of acronyms, with each new way of doing things trying to outperform the others on benchmark tasks. [image: Multiple acronyms in the background with a bar graph labeled "% classification error" on the y-axis and various acronyms labeled on the x-axis]

Despite all this variety, every neural net contains the same two common building blocks: There's a part that's linear, which is usually code for "things work out nicely." [image: linear graph illustration] Then there's a part that's nonlinear, which often means "lol good luck trying to prove anything." [image: nonlinear graph illustration depicting a dragon over a squiggly line]

The linear part is that synthesis step. In a neural net, there are a bunch of nodes (or neurons) that are connected by edges with different weights. [image: multiple equations being merged together with arrows being pointed at a blue dot that is generating the solution] Each node adds up the input from all its neighbors, multiplied by the weight of the edges connecting them, to get a number representing its state.

The nonlinear part is the decision step. Based on the number from the linear step, the node decides: Blips? No blips?? [image: confused blue dot with thought bubbles including blip or no blip]

Here's a cool thing: For as scary and annoying as nonlinearities can be, the ones that seem to work best for neural nets are all... pretty nice. [image: three graphs; one with negative values going to 0 and positive values going to 1, another being same as the last one, but smoother, and the last one having negative values going to 0 and positive values staying the same]

Which is kind of amazing. [image: "Step 1. Adding & Multiplying" with arrows pointing at a dot "Step 2. A Simple Nonlinearity" with a dot and a nonlinear graph] Neural nets have crushed the competition in tons of machine learning challenges, yet the basic rules for their nodes are simple and clean. It's the fact that they're combined into networks that give rise to the bogglingly good performance.

What do these networks look like? There are input neurons, that ingest inputs. Like the third pixel from the right and five down in an image. [image: cat picture with pixels that have arrows pointing at input neurons] There are output neurons, that give outputs. There might be ten of these if you're trying to classify the digits 0 through 9, or a LOT more if you're trying to identify the content of an image. [image: output neurons saying "cat", "shoe", and "bean"]

There are typically many more neurons in between the inputs and outputs, organized into layers. [image: network graph] They talk to their neighbors by alternating between simple linear steps and (still pretty simple) nonlinear steps.

Before training, the outputs from the output layer will be nonsense. [image: neural dots outputting scribbled musical notes] They won't match the data the same way a song played on a randomly tuned piano won't sound anything like what it's supposed to.

Training the network is like tuning that piano. [image: neural dots pointing to a musical note that gradually becomes more defined] Just like you'd tighten or loosen strings in a piano to make a song sound right, you increase or decrease the edge weights in a neural net to match its outputs to the data you're training on.

By tuning the edge weights like this during training, the network is primed to handle the same kind of inputs during testing. [image: network graph with a chat bubble saying "cat"]

There are a lot of flavors of neural net right now, and we don't really know which approaches are going to win in the long-run. [image: network graphs with legs running towards the finish line of a race] In fact, there's a lot we don't know about why neural nets work... at all...

But we're learning. And we know enough now to make informed choices about which tools to use for which problems. [image: woman in profile looking up]

We're learning that convolutional layers are good for exploiting the spatial structure of images. [image: picture of a dog being analyzed] We're learning that randomly tossing out parts of the data at each iteration can make the fit more robust. [image: picture of a cat followed by a duplicate image with an ear blacked out] We're learning that one of the simplest of the simple nonlinear activation functions, ReLu, seems to work best in deep learning. [image: nonlinear graph with max (x,0)]

If you're new to neural nets, how do you start forming insights like this of your own? [image: man thinking]

You get your hands on some data... [image: man with a box full of cards with numbers on them] ... and you start playing around. [image: hand dumps the box over]

You can tackle the first part with The Wolfram Data Repository, where lots of different datasets are already available and pre-processed for your use. You can grab a dataset and download it into your notebook with a line of code. [image: a screenshot of a sample dataset from the Wolfram Data Repository with handwritten digit dataset and pre-cleaned training & test sets]

As for playing around with it? The Wolfram Language's neural net framework makes that straightforward as well. [image: a screenshot of a pre-built neural net from the Wolfram Neural Net Repository] You can download a pre-built neural net from the Wolfram Neural Net Repository that's been trained to match the data, then you can look at its guts.

The Neural Net Repository gives you a lot of options. [image: hand reaching for one of three jars that has network graphs in them] Audio? Text? Images? No problem. You can pick a net that's a good fit for your problem right off the shelf with NetModel and tune it to match your data.

Or you can start from scratch: create your own networks using NetChain, train them, and see how performance changes as you knock layers out or add them in. [image: a screenshot of creating networks using NetChain with text highlighting how to create net, train it, and check its performance]

This kind of self-guided exploration can help you go from being familiar with the idea of a neural network to developing an intuition for them. [image: man on a computer with a lightbulb moment] And you can get up and running on all sorts of data types in minutes by tapping into the curated datasets the Wolfram Language provides.

But this is a lot more than a sandbox. The real power of the framework is how rapidly you can put its tools to work on your own datasets... [image: signal lines with network graph illustration]

... And how easily you can share your work with others. [image: woman handing a red box labeled "box contains USB with a .NB on it" over to a man]

With only a few lines of code, you can begin tackling problems of your own with the tools in the Wolfram Language. [image: red dot network graph illustration] Get started today with Wolfram|One or Wolfram Mathematica.

[swimmer963]

Ganbreeder is the beeeest

So I think I have discovered the thing that has the effect on my brain that addictive phone flash games do for many people, and it is ganbreeder.app [generative adversarial networks for image generation]. “Ganbreeder is a collaborative art tool for discovering images. Images are 'bred' by having children, mixing with other images and being shared via their URL.” 

You can see my saved images here: https://ganbreeder.app/mirandadixon

Some thoughts: 

-This has the same qualia as my hypnagogic imagery, which may be why it’s so goddamned addictive. I spent hours on it yesterday and then had very vivid dreams. 

-It’s *fascinating* to see which points in imagespace are adjacent, and which ones are “sticky” / attractors / easy to converge on. I can hang out for a while in the space of “weird fucked-up birds” (https://ganbreeder.app/i?k=ed48f5ca7a91cd060c9187d6) or “extremely round furry animals” (https://ganbreeder.app/i?k=180ffd4c2809af10b6cd0a64) or “spiky things that look like bacteriophage-robots” (https://ganbreeder.app/i?k=925a4fd8bf9715be7f85b057) or “spider-clocks”. 

-Some of the building-landscapes are *beautiful* and gloriously evocative (https://ganbreeder.app/i?k=52c99530cba54fcc26fccd33). A lot of them look like they are on fire (https://ganbreeder.app/i?k=607d2c4286a59fa261827782). 

-Whatever points generate round balls in the center field are weirdly unstable - the “children” are almost always different. Abstract images are also fairly unstable, and often try to turn into landscapes or closeups of sea life. 

-The networks are good at landscapes and buildings, and bad at clothing or people (https://ganbreeder.app/i?k=20f2d3071e080aea9b5139ba). They’re bizarrely good at dogs and terrible at snakes. 

-I can fairly reliably get from birds to buildings via the route of things-that-look-like-deformed-musical-instruments. 

-Any writing it attempts inevitably looks like the symbols for some arcane ritual. (Ritual stone: https://ganbreeder.app/i?k=e0a848daea8083d1327603a2, ritual chest: https://ganbreeder.app/i?k=36f59f2f0f8cfad8263b5c27) 

-I produced this alternate-universe bottle of hand sanitizer (https://ganbreeder.app/i?k=4897a81ff4bb8a95be2a5fd50) by breeding a radiator with nipples. 

-I think the “volcano dog” (https://ganbreeder.app/i?k=378589c190c5476468da5d18) is one of my all-time favourites. 

-Tarantula cat! https://ganbreeder.app/i?k=1dec3d39a8c29afbe8ff2222

-Most of the points adjacent to close-ups of food look *really* gross. 

-It’s surprisingly possible to get images of things that look like they’re painted in a watercolour style. I’m not sure if this was true for any of the training data.