Starting this week, I'll be doing a new series called Deep Learning Research Review. Every couple weeks or so, I'll be summarizing and explaining research papers in specific subfields of deep learning. This week I'll begin with Generative Adversarial Networks. According to Yann LeCun, "adversarial training is the coolest thing since sliced bread". I'm inclined to believe so because I don't think sliced bread ever created this much buzz and excitement within the deep learning community.

Update 11/29: I'm looking for translators to help translate this post into different languages, particularly Chinese (中文), Spanish (Español), Korean (한국어), Russian (ру сский язы к), and Japanese (日本語). When I read Machine Learning papers, I ask myself whether the contributions of the paper fall under improvements to 1) Expressivity 2) Trainability, and/or 3) Generalization. I learned this categorization from my colleague Jascha Sohl-Dickstein at Google Brain, and the terminology is also introduced in this paper. I have found this categorization effective in thinking about how individual research papers (especially on the theoretical side) tie subfields of AI research (e.g. In this blog post, I discuss how these concepts tie into current (Nov 2017) machine learning research on Supervised Learning, Unsupervised Learning, and Reinforcement Learning. I consider Generalization to be comprised of two categories -- "weak" and "strong" generalization -- and I will discuss them separately.

Artificial intelligence (AI) can obtain unbelievably accurate insights into a neighborhood's inhabitants – from their income and level of education to their ethnic background and political beliefs – just by looking at images from Google Street View. If, for example, you wanted to see whether an area voted Republican or Democrat, the AI algorithm would be able to correctly tell you with over 80 percent accuracy, namely based on the types of vehicles riding on the road. The deep-learning algorithm was developed by a team of computer scientists based at Stanford University. Their study was published in the Proceedings of the National Academy of Sciences. Throughout this process, it used an object recognition algorithm to clock tens of millions of houses, landscape features like shrubberies, and – most importantly – vehicles.

In this work we propose a method for anatomical data augmentation that is based on using slices of computed tomography (CT) examinations that are adjacent to labeled slices as another resource of labeled data for training the network. The extended labeled data is used to train a U-net network for a pixel-wise classification into different hepatic lesions and normal liver tissues. Our dataset contains CT examinations from 140 patients with 333 CT images annotated by an expert radiologist. We tested our approach and compared it to the conventional training process. Results indicate superiority of our method. Using the anatomical data augmentation we achieved an improvement of 3% in the success rate, 5% in the classification accuracy, and 4% in Dice.

We propose denoising dictionary learning (DDL), a simple yet effective technique as a protection measure against adversarial perturbations. We examined denoising dictionary learning on MNIST and CIFAR10 perturbed under two different perturbation techniques, fast gradient sign (FGSM) and ja-cobian saliency maps (JSMA). We evaluated it against five different deep neural networks (DNN) representing the building blocks of most recent architectures indicating a successive progression of model complexity of each other. We show that each model tends to capture different representations based on their architecture. For each model we recorded its accuracy both on the perturbed test data previously misclassified with high confidence and on the denoised one after the reconstruction using dictionary learning. The reconstruction quality of each data point is assessed by means of PSNR (Peak Signal to Noise Ratio) and Structure Similarity Index (SSI). We show that after applying (DDL) the reconstruction of the original data point from a noisy sample results in a correct prediction with high confidence.

Sharone Dayan MVA ENS Paris-Saclay sharone.dayan@ens-paris-saclay.fr Manual segmentation of the Left Ventricle (LV) is a tedious and meticulous task that can vary depending on the patient, the Magnetic Resonance Images (MRI) cuts and the experts. Still today, we consider manual delineation done by experts as being the ground truth for cardiac diagnosticians. Thus, we are reviewing the paper - written by Avendi and al. - who presents a combined approach with Convolutional Neural Networks, Stacked Auto-Encoders and Deformable Models, to try and automate the segmentation while performing more accurately. Furthermore, we have implemented parts of the paper (around three quarts) and experimented both the original method and slightly modified versions when changing the architecture and the parameters.

Rapid advances of hardware-based technologies during the past decades have opened up new possibilities for Life scientists to gather multimodal data in various application domains (e.g., Omics, Bioimaging, Medical Imaging, and [Brain/Body]-Machine Interfaces), thus generating novel opportunities for development of dedicated data intensive machine learning techniques. Overall, recent research in Deep learning (DL), Reinforcement learning (RL), and their combination (Deep RL) promise to revolutionize Artificial Intelligence. The growth in computational power accompanied by faster and increased data storage and declining computing costs have already allowed scientists in various fields to apply these techniques on datasets that were previously intractable for their size and complexity. This review article provides a comprehensive survey on the application of DL, RL, and Deep RL techniques in mining Biological data. In addition, we compare performances of DL techniques when applied to different datasets across various application domains. Finally, we outline open issues in this challenging research area and discuss future development perspectives.

We consider the problem of training generative models with deep neural networks as generators, i.e. to map latent codes to data points. Whereas the dominant paradigm combines simple priors over codes with complex deterministic models, we argue that it might be advantageous to use more flexible code distributions. We demonstrate how these distributions can be induced directly from the data. The benefits include: more powerful generative models, better modeling of latent structure and explicit control of the degree of generalization.

Summary: Quantum computing is now a commercial reality. Here's the story of the companies that are currently using it in operations and how this will soon disrupt artificial intelligence and deep learning. Like a magician distracting us with one hand while pulling a fast one with the other Quantum computing has crossed over from research to commercialization almost without us noticing. Has the dream of Quantum computing actually stepped out of the lab into the world of actual application. Well, Lockheed Martin has been using it for seven years.

Recently Google DeepMind announced AlphaGo Zero -- an extraordinary achievement that has shown how it is possible to train an agent to a superhuman level in the highly complex and challenging domain of Go, 'tabula rasa' -- that is, from a blank slate, with no human expert play used as training data. It thrashed the previous reincarnation 100–0, using only 4TPUs instead of 48TPUs and a single neural network instead of two. Click on the image to zoom in. To read more and access the full cheat sheet, click here.