Basically, they create a policy network, which is a convolutional neural network, that predicts the next move a human player would do from a board state. They create a value network, also a convolutional neural network, that predicts the outcome (win or lose) of the game given the current board state. Continue reading “Mastering the Game of Go – slides [paper explained]”
In this work, we used a convolutional neural network (CNN) to classify 10 different types of skin lesions, including melanoma and non-melanoma types.
The key technical contribution was to use multiple tracts (or paths) within the neural network, to train (and test) the network on an image using multiple image resolutions simultaneously. Additionally, we extended a CNN pretrained on a single image resolution to work for multiple image resolutions.
There’s this really neat new idea on how to train neural networks that recently came out know as generative adversarial nets (GAN).
The basic idea of a GAN is to train two networks to compete with each other (hence the name “adversarial“). One network (called the generator) creates images that look just like real images. The other network (called the discriminator) distinguishes between real images and those images the generator produced.
Thus the two networks compete with each other, where the generator generates images to fool the discriminator, and the discriminator discriminates between the generator’s images and real images.
Here is a combined short summary on my travels to the city of Prague in the Czech Republic along with corresponding images created using Google’s DeepDreams.
What is this DeepDreams you speak of?
Basically, DeepDream is a deep neural network that was trained to recognize objects from millions of images. A deep neural network is composed of a stack of layers. Basically, these layers learn image filters that when applied to an image classify the image (e.g., is this an image of a cat or a dog?).
You give DeepDream an image and specify a layer in the neural network. The original image is then slightly perturbed to create a modified image that causes the specified layer in the neural network to be more activated.
Early layers in the neural network are sensitive to low level concepts like the edges and textures in the image. So if you specify an early layer, your image will be modified to have edges and textures that most activate the early selected layer.
Later (or deeper) layers in the neural network are activated when they see higher level concepts such as faces. So any areas in the original image that slightly look like a face, will be modified to look more like a face.
Okay, but now you might ask, but what about Prague? How was your trip? Did you like the city?
Yeah it was nice! Thanks for asking. Did you want to see some pictures? Here’s one of an old building.
Let’s try some deep dreaming on this. We’ll use the neural network known as VGG16 (it’s a famous neural network that performs very well on competitions). We’ll start by telling VGG16 (the neural network) to modify this image so that one of it’s middle layers becomes more activated. Specifically, we will activate layerconv3_1 from VGG16 (if you don’t know what conv3_1 means, that’s okay – it’s just a technical detail specifying what layer to use). This gives us this:
Now if we activate a deeper layer, conv5_2, we get this crazy looking image,
In this work, we looked at how to classify skin lesions from images captured with a digital camera (i.e., non-dermoscopy). Our approach was able to distinguish among 10 different types of skin diseases over 1300 images and achieved an accuracy higher than what was previously reported over the same dataset. We did this by applying deep learning (i.e., pretrained convolutional neural networks) to melanoma and non-melanoma skin images.
Let’s summarize it in two lines: The authors proposed a way to combine information from an image and a corresponding text caption. They use a Recurrent Neural Network (RNN) to then generate text captions that describe the image.
Pretty interesting stuff. This is the first paper where I really took a close look at RNNs as well.