The loss functions of deep neural networks are complex and their geometric properties are not well understood. We show that the optima of these complex loss functions are in fact connected by a simple polygonal chain with only one bend, over which training and test accuracy are nearly constant. We introduce a training procedure to discover these high-accuracy pathways between modes. Inspired by this new geometric insight, we propose a new ensembling method entitled Fast Geometric Ensembling (FGE). Using FGE we can train high-performing ensembles in the time required to train a single model. We achieve improved performance compared to the recent state-of-the-art Snapshot Ensembles, on CIFAR-10 and CIFAR-100, using state-of-the-art deep residual networks. On ImageNet we improve the top-1 error-rate of a pre-trained ResNet by 0.56% by running FGE for just 5 epochs.

We combine multi-task learning and semi-supervised learning by inducing a joint embedding space between disparate label spaces and learning transfer functions between label embeddings, enabling us to jointly leverage unlabelled data and auxiliary, annotated datasets. We evaluate our approach on a variety of sequence classification tasks with disparate label spaces. We outperform strong single and multi-task baselines and achieve a new state-of-the-art for aspect- and topic-based sentiment analysis.

Along with the advance of opinion mining techniques, public mood has been found to be a key element for stock market prediction. However, in what manner the market participants are affected by public mood has been rarely discussed. As a result, there has been little progress in leveraging public mood for the asset allocation problem, as the application is preferred in a trusted and interpretable way. In order to address the issue of incorporating public mood analyzed from social media, we propose to formalize it into market views that can be integrated into the modern portfolio theory. In this framework, the optimal market views will maximize returns in each period with a Bayesian asset allocation model. We train two neural models to generate the market views, and benchmark the performance of our model using market views on other popular asset allocation strategies. Our experimental results suggest that the formalization of market views significantly increases the profitability (5% to 10%) of the simulated portfolio at a given risk level.

In this work, we focus on multilingual systems based on recurrent neural networks (RNNs), trained using the Connectionist Temporal Classification (CTC) loss function. Using a multilingual set of acoustic units poses difficulties. To address this issue, we proposed Language Feature Vectors (LFVs) to train language adaptive multilingual systems. Language adaptation, in contrast to speaker adaptation, needs to be applied not only on the feature level, but also to deeper layers of the network. In this work, we therefore extended our previous approach by introducing a novel technique which we call "modulation". Based on this method, we modulated the hidden layers of RNNs using LFVs. We evaluated this approach in both full and low resource conditions, as well as for grapheme and phone based systems. Lower error rates throughout the different conditions could be achieved by the use of the modulation.

Supermarket cashiers and self-scan checkouts could become redundant thanks to a new app that speeds up the weekly food shop. The app, called Supersmart, allows users to scan items as they shop using their smartphone and then wheel them to a special floor pad. Several cameras and a deep-learning algorithm combine to detect all the items, before the customer pays for them at an automated checkout. It is hoped the technology will be quicker than both self-scan checkouts and human-operated tills. Supermarket cashiers and self-scan checkouts could become redundant thanks to a new app that speeds up the weekly food shop.

It's still relatively rare for artificial intelligence to deliver a crushing victory over human physicians in a head-to-head test of medical expertise. But a deep neural network approach managed to beat 42 dermatology experts in diagnosing a common nail fungus that affects about 35 million Americans each year. The latest successful demonstration of AI's capabilities in the medical field relied heavily upon a team of South Korean researchers putting together a huge dataset of almost 50,000 images of toenails and fingernails. That large amount of data used to train the deep neural networks on recognizing cases of onychomycosis--a common fungal infection that can make nails discolored and brittle--provided the crucial edge that enabled deep learning to outperform medical experts. "This study was the first to show that AI has overwhelmed the specialists," says Seung Seog Han, a dermatologist and clinician at I Dermatology in Seoul, South Korea.

Secret to GoogleBrain and RankBrain algorithm revealed. One is going to give a historical overview about GoogleBrain and analyse the pattern, then we will conclude our finding about the current situation and future changes in search engine algorithm. Back in 2006 there were some interests in implementing artificial intelligence in Google search engine algorithm. A few years later in 2014, GoogleBrain was established after acquisition of DeepMind, a British artificial intelligence company which was founded in 2010. They worked on how to play video games based on machine learning and artificial neural networks (ANNs).

Being able to learn from mistakes is a powerful ability that humans (being mistake-prone) take advantage of all the time. Even if we screw something up that we're trying to do, we probably got parts of it at least a little bit correct, and we can build off of the things that we did not to do better next time. Robots can use similar trial-and-error techniques to learn new tasks. With reinforcement learning, a robot tries different ways of doing a thing, and gets rewarded whenever an attempt helps it to get closer to the goal. Based on the reinforcement provided by that reward, the robot tries more of those same sorts of things until it succeeds. Where humans differ is in how we're able to learn from our failures as well as our successes.

Last week I published a brief analysis of the OpenMined platform as one of the new technologies that is trying to enable truly decentralized artificial intelligence(AI) processes by leveraging blockchain technologies. In the article, I mentioned that OpenMined drew parts of its inspiration from Google's research about federated learning as a mechanism to improve on the traditional centralized approach to train AI models. From my perspective, I consider federated learning is one of the most interesting AI research breakthroughs of the last two years that is already powering mission critical applications. The idea behind federated learning is as conceptually simple as it its technologically complex. Traditional machine learning programs relied on a centralized model for training in which a group of servers run a specific model against training and validation datasets.

Alphabet's artificial intelligence arm DeepMind and the U.S. Department of Veterans Affairs have unveiled a research partnership focused on predicting patient deterioration in the hospital setting. The issue is the cause of approximately 11 percent of in-hospital deaths, NHS research shows. Together, the organizations will examine 700,000 historical, depersonalized patient medical records. They'll analyze patterns from the data to see if machine learning can pinpoint risk factors for patient deterioration. To start, the relationship will zoom in on acute kidney injury, a complication related to patient deterioration.