52

First up, those images (even the first few) aren't complete trash despite being junk to humans; they're actually finely tuned with various advanced techniques, including another neural network. The deep neural network is the pre-trained network modeled on AlexNet provided by Caffe. To evolve images, both the directly encoded and indirectly encoded images, ...


26

Yes, there is some research on this topic, which can be called adversarial machine learning, which is more an experimental field. An adversarial example is an input similar to the ones used to train the model, but that leads the model to produce an unexpected outcome. For example, consider an artificial neural network (ANN) trained to distinguish between ...


25

The images that you provided may be unrecognizable for us. They are actually the images that we recognize but evolved using the Sferes evolutionary framework. While these images are almost impossible for humans to label with anything but abstract arts, the Deep Neural Network will label them to be familiar objects with 99.99% confidence. This result ...


14

Now the question is how to handle such large image sizes where there is no privileges of downsampling I assume that by downsampling you mean scaling down the input before passing it into CNN. Convolutional layer allows to downsample the image within a network, by picking a large stride, which is going to save resources for the next layers. In fact, that's ...


13

All answers here are great, but, for some reason, nothing has been said so far on why this effect should not surprise you. I'll fill the blank. Let me start with one requirement that is absolutely essential for this to work: the attacker must know neural network architecture (number of layers, size of each layer, etc). Moreover, in all cases that I examined ...


13

In a convolutional neural network, is there a unique filter for each input channel or are the same new filters used across all input channels? The former. In fact there is a separate kernel defined for each input channel / output channel combination. Typically for a CNN architecture, in a single filter as described by your number_of_filters parameter, ...


12

The following picture that you used in your question, very accurately describes what is happening. Remember that each element of the 3D filter (grey cube) is made up of a different value (3x3x3=27 values). So, three different 2D filters of size 3x3 can be concatenated to form this one 3D filter of size 3x3x3. The 3x3x3 RGB chunk from the picture is ...


12

Sometimes if the rules used by an AI to identify characters are discovered, and if the rules used by a human being to identify the same characters are different, it is possible to design characters that are recognized by a human being but not recognized by an AI. However, if the human being and AI both use the same rules, they will recognize the same ...


10

An important question that does not yet have a satisfactory answer in neural network research is how DNNs come up with the predictions they offer. DNNs effectively work (though not exactly) by matching patches in the images to a "dictionary" of patches, one stored in each neuron (see the youtube cat paper). Thus, it may not have a high level view of the ...


10

Convolutional Nets (CNN) rely on mathematical convolution (e.g. 2D or 3D convolutions), which is commonly used for signal processing. Images are a type of signal, and convolution can equally be used on sound, vibrations, etc. So, in principle, CNNs can find applications to any signal, and probably more. In practice, there exists already work on NLP (as ...


10

Yes there are, for instance one pixel attacks described in Su, J.; Vargas, D.V.; Kouichi, S. One pixel attack for fooling deep neural networks. arXiv:1710.08864 One pixels attacks are attacks in which changing one pixel in input image can strongly affect the results.


7

The neural network is typically a set size once it's created. You'd have to create a network big enough for your data-set.


7

The white noise that fools DNNs isn't really white noise. It has been altered in the same way as the synthetic misclassified pictures have been altered. You have to change many input pixels in exactly such a way, that these little changes aren't perceptible, but propagated through the network add up to a misclassification. This is not going to happen by ...


6

Well, I do not know what type of features you are giving to your neural network. However, in general, I would go with a single neural network. It seems that you have no limitation in resources for training your network and the only problem is resources while you apply your network. The thing is that probably the two problems have things in common (e.g. ...


6

I believe they don't use CNNs. The most important reason why it's because they have more information than a regular image: time. The input they receive is a sequence of (x,y,t) as you draw on the screen, which they refer as "ink". This gives them the construction of the image for free, which a CNN would have to deduce by itself. They tried two approaches. ...


6

Yes, this is possible. There is actually a pretty easy way that doesn't even require machine learning and can be implemented with a small amount of code. You just use a framework for image processing (e.g. PIL for Python), find the marks by going over your image with an appropriate filter and use the implemented crop function, that the framework hopefully ...


5

In May 2016 Google announced a custom ASIC which was is specifically built for machine learningwiki and tailored for TensorFlow. It is using tensor processing unit (TPU) which is a programmable microprocessor designed to accelerate artificial neural networks. NeuroCores, 12x14 sq-mm chips which can be interconnected in a binary tree, see: Neurogrid, a ...


5

How is it possible that deep neural networks are so easily fooled? Deep neural networks are easily fooled by giving high confidence predictions for unrecognizable images. How is this possible? Can you please explain ideally in plain English? Intuitively, extra hidden layers ought to make the network able to learn more complex classification functions, ...


5

Yes. Here are some of the most prominent ones and their respective state-of-the-art errors: CIFAR-10: ~3.5% error CIFAR-100: ~24% error STL-10: ~26% error SVHN: ~1.7% error ImageNet tasks: the best 2012 classification task solution got 15% top-5 error, better results are currently available You can check an updated list of solutions here. Also, a more ...


5

It's an interesting question about what makes humans unique. There is a good book on the subject titled What Computers Cant Do by Hubert Dreyfus. One task that a computer can't handle (for now at least) is ranking important things. For example, CAPTCHA asks you to order a random list of things (small one, five or six items) by importance. This particular ...


5

A method that could possibly work is utilising optical illusions such as one where two lines down a hallway are identical but one seems longer to the human eye, then they could be prompted with a multiple choice question as to the state of the line, which to our eyes looks longer, but to a computer, is still the same length of line. Of course, there is ...


5

There are several approaches as to how this can be achieved. One recent study from 2015 about Action Recognition in Realistic Sports VideosPDF uses the action recognition framework based on the three main steps of feature extraction (shape, post or contextual information), dictionary learning to represent a video, and classification (BoW framework). A few ...


5

This study from 2012 uses 3D convolutional neural networks (CNN) for automated recognition of human actions in surveillance videos. The 3D CNN model extracts features from both the spatial and the temporal dimensions by performing 3D convolutions, thereby capturing the motion information encoded in multiple adjacent frames. A very similar deep learning ...


5

The simple answer is "no, they aren't limited to images": CNNs are also being used for natural language processing. (See here for an introduction.) I haven't seen them applied to graphical data yet, but I haven't looked; there are some obvious things to try and so I'm optimistic that it would work.


5

We are getting pretty good at image generation, some examples: Radford, Alec, Luke Metz, and Soumith Chintala. "Unsupervised representation learning with deep convolutional generative adversarial networks." arXiv preprint arXiv:1511.06434 (2015). https://arxiv.org/pdf/1511.06434.pdf Gregor, Karol, Ivo Danihelka, Alex Graves, Danilo Jimenez Rezende, and Daan ...


5

Usually for images the feature set is the pixel density values and in this case it will lead to quite a big feature set; also down sampling the images is also not recommended as you may lose (actually will) loose important data. [1] But there are some techniques that can help you reduce the feature set size, approaches like PCA(Principle Component Analysis) ...


5

Here's an example: How to hack your face to dodge the rise of facial recognition tech In his recent book The Fall, Stephenson wrote about smartglasses that that project a pattern over the facial features to foil recognition algorithms (which seems not only feasible but likely;) Here's an article from our sponsors, Adversarial AI: As New Attack Vector ...


4

One of the Pinterest's white paper about Human Curation and Convnets powering item-to-item recommendationsarxiv describes implementation of convolutional neural network (CNN) based visual features (VGG2014, Faster R-CNN). This demonstrates the effectiveness of it (such image or object representations) which can improve user engagement. The visual features ...


4

As the name implies, algorithmic bias is related with the used algorithm. Due to the way it was programmed or devised, the algorithm will be biased in some of its samples. From Communications of the ACM: [Algorithms] often inadvertently pick up the human biases that are incorporated when the algorithm is programmed, or when humans interact with that ...


4

First you should give it a try, because anyone's guess could be off, as there isn't really a complete high level analytic model of how neural networks behave on real data. Most results with neural networks are informed by theory, but involve a whole lot of experimental testing. I suspect your network might at least learn something from the new images, but ...


Only top voted, non community-wiki answers of a minimum length are eligible