Each of Burger King's new ads starts with an anachronistic burst of noise from a dial-up modem and a solemn warning: "This ad was created by artificial intelligence." Then, over shots of glistening burgers and balletic fries, a robotic-sounding narrator deploys exactly the sort of clunky grammar and conceptual malapropisms we expect from a dumb AI. "The chicken crossed the road to become a sandwich. Burger King encouraged the chicken," says the voice. In a press release, Burger King claims the videos are the work of a "new deep learning algorithm," but an article from AdAge makes it clear that humans -- not machines -- are responsible for the funnies. "Artificial intelligence is not a substitute for a great creative idea coming from a real person," Burger King's global head of brand marketing, Marcelo Pascoa, told the publication.

Microsoft Ignite, Microsoft's annual developer conference, wrapped up last week and many of the big announcements focused on artificial intelligence and machine learning. The keynote presentation from Microsoft's Cloud AI lead Eric Boyd showcases the major developments, or you can check out his accompanying blog post for a written summary. In this post, I'll dive down into the details of the specific announcements and provide links to the associated recorded presentations from Ignite and other technical documents. Azure Machine Learning Service (Azure ML) adds several new capabilities. These include a new Python package that developers can use to prepare data for analysis, automate machine learning and deep learning model training, track model performance under various conditions, and automate the model selection process (see below).

Smart manufacturing refers to using advanced data analytics to complement physical science for improving system performance and decision making. With the widespread deployment of sensors and Internet of Things, there is an increasing need of handling big manufacturing data characterized by high volume, high velocity, and high variety. Deep learning provides advanced analytics tools for processing and analysing big manufacturing data. This paper presents a comprehensive survey of commonly used deep learning algorithms and discusses their applications toward making manufacturing "smart". The evolvement of deep learning technologies and their advantages over traditional machine learning are firstly discussed.

This past Tuesday, Google and Facebook announced a partnership to enable the open-sourced machine learning framework PyTorch to work with Tensor-Processing Units (TPUs). This partnership could signal a new age of collaboration towards AI research. "Today, we're pleased to announce that engineers on Google's TPU team are actively collaborating with core PyTorch developers to connect PyTorch to Cloud TPUs. The long-term goal is to enable everyone to enjoy the simplicity and flexibility of PyTorch while benefiting from the performance, scalability, and cost-efficiency of Cloud TPUs." - Director of Product, Rajen Sheth PyTorch is Facebook's open-source framework which enables development of mathematical programs like those used in Artificial Intelligence research. Such frameworks allow researchers to develop arbitrarily complicated mathematical computational graphs and automatically calculate derivatives.

Learning from data has led to a paradigm shift in computational materials science. In particular, it has been shown that neural networks can learn the potential energy surface and interatomic forces through examples, thus bypassing the computationally expensive density functional theory calculations. Combining many-body techniques with a deep learning approach, we demonstrate that a fully-connected neural network is able to learn the complex collective behavior of electrons in strongly correlated systems. Specifically, we consider the Anderson-Hubbard (AH) model, which is a canonical system for studying the interplay between electron correlation and strong localization. The ground states of the AH model on a square lattice are obtained using the real-space Gutzwiller method. The obtained solutions are used to train a multi-task multi-layer neural network, which subsequently can accurately predict quantities such as the local probability of double occupation and the quasiparticle weight, given the disorder potential in the neighborhood as the input.

Mobile electrocardiogram (ECG) recording technologies represent a promising tool to fight the ongoing epidemic of cardiovascular diseases, which are responsible for more deaths globally than any other cause. While the ability to monitor one's heart activity at any time in any place is a crucial advantage of such technologies, it is also the cause of a drawback: signal noise due to environmental factors can render the ECGs illegible. In this work, we develop convolutional neural networks (CNNs) to automatically label ECGs for noise, training them on a novel noise-annotated dataset. By reducing distraction from noisy intervals of signals, such networks have the potential to increase the accuracy of models for the detection of atrial fibrillation, long QT syndrome, and other cardiovascular conditions. Comparing several architectures, we find that a 16-layer CNN adapted from the VGG16 network which generates one prediction per second on a 10-second input performs exceptionally well on this task, with an AUC of 0.977.

We study the topology of the space of learning tasks, which is critical to understanding transfer learning whereby a model such as a deep neural network is pre-trained on a task, and then used on a different one after some fine-tuning. First we show that using the Kolmogorov structure function we can define a distance between tasks, which is independent on any particular model used and, empirically, correlates with the semantic similarity between tasks. Then, using a path integral approximation, we show that this plays a central role in the learning dynamics of Deep Networks, and in particular in the reachability of one task from another. We show that the probability of paths connecting two tasks, is asymmetric and has a static component that depends on the geometry of the loss function, in particular on the curvature, and a dynamic component that is model dependent and relates to the ease of traversing such paths. Surprisingly, the static component corresponds to the distance derived from the Kolmogorov Structure Function. With the dynamic component, this gives strict lower bounds on the complexity necessary to learn a task starting from the solution to another. Our analysis also explains more complex phenomena where semantically similar tasks may be unreachable from one another, a phenomenon called Information Plasticity and observed in diverse learning systems such as animals and deep neural networks.

We marry two powerful ideas: deep representation learning for visual recognition and language understanding, and symbolic program execution for reasoning. Our neural-symbolic visual question answering (NS-VQA) system first recovers a structural scene representation from the image and a program trace from the question. It then executes the program on the scene representation to obtain an answer. Incorporating symbolic structure as prior knowledge offers three unique advantages. First, executing programs on a symbolic space is more robust to long program traces; our model can solve complex reasoning tasks better, achieving an accuracy of 99.8% on the CLEVR dataset. Second, the model is more data- and memory-efficient: it performs well after learning on a small number of training data; it can also encode an image into a compact representation, requiring less storage than existing methods for offline question answering. Third, symbolic program execution offers full transparency to the reasoning process; we are thus able to interpret and diagnose each execution step.

Based on MacKay's information theoretic model of supervised machine learning [23], this article discusses how to practically estimate the maximum size of a neural network given a training data set. First, we present four easily applicable rules to analytically determine the capacity of neural network architectures. This allows the comparison of the efficiency of different network architectures independently of a task. Second, we introduce and experimentally validate a heuristic method to estimate the neural network capacity requirement for a given dataset and labeling. This allows an estimate of the required size of a neural network for a given problem. We conclude the article with a discussion on the consequences of sizing the network wrongly, which includes both increased computation effort for training as well as reduced generalization capability.

Rewards are sparse in the real world and most today's reinforcement learning algorithms struggle with such sparsity. One solution to this problem is to allow the agent to create rewards for itself -- thus making rewards dense and more suitable for learning. In particular, inspired by curious behaviour in animals, observing something novel could be rewarded with a bonus. Such bonus is summed up with the real task reward -- making it possible for RL algorithms to learn from the combined reward. We propose a new curiosity method which uses episodic memory to form the novelty bonus. To determine the bonus, the current observation is compared with the observations in memory. Crucially, the comparison is done based on how many environment steps it takes to reach the current observation from those in memory -- which incorporates rich information about environment dynamics. This allows us to overcome the known "couch-potato" issues of prior work -- when the agent finds a way to instantly gratify itself by exploiting actions which lead to unpredictable consequences. We test our approach in visually rich 3D environments in VizDoomand DMLab. In VizDoom, our agent learns to successfully navigate to a distant goal at least 2 times faster than the state-of-the-art curiosity method ICM. In DMLab, our agent generalizes well to new procedurally generated levels of the game -- reaching the goal at least 2 times more frequently than ICM on test mazes with very sparse reward. Many real-world tasks have sparse rewards. For example, animals searching for food may need to go many miles without any reward from the environment. Multiple approaches were proposed to achieve better explorative policies. One way is to give a reward bonus which facilitates exploration by rewarding novel observations. The reward bonus is summed up with the original task reward and optimized by standard RL algorithms. Such an approach is motivated by neuroscience studies of animals: an animal has an ability to reward itself for something novel - the mechanism biologically built into its dopamine release system.