We've heard from customers that scaling TensorFlow training jobs to multiple nodes and GPUs successfully is hard. TensorFlow has distributed training built-in, but it can be difficult to use. Recently, we made optimizations to TensorFlow and Horovod to help AWS customers scale TensorFlow training jobs to multiple nodes and GPUs. With these improvements, any AWS customer can use an AWS Deep Learning AMI to train ResNet-50 on ImageNet in just under 15 minutes. To achieve this, 32 Amazon EC2 instances, each with 8 GPUs, a total 256 GPUs, were harnessed with TensorFlow. All of the required software and tools for this solution ship with the latest Deep Learning AMIs (DLAMIs), so you can try it out yourself. You can train faster, implement your models faster, and get results faster than ever before. This blog post describes our results and shows you how to try out this easier and faster way to run distributed training with TensorFlow. Figure A. ResNet-50 ImageNet model training with the latest optimized TensorFlow with Horovod on a Deep Learning AMI takes 15 minutes on 256 GPUs.

It's difficult to think of 2018 as a year with anything worth celebrating. But despite all the bad news the year dealt us, there were successes -- if you know where to look. In all corners of tech, we saw wins big and small. There were advances in obvious categories like laptops, smartphones and the connected home, but we also looked outside the mainstream for some of the more surprising gems. Think mini synthesizers for music nerds, retro emulators for nostalgic gamers and e-readers for modern book snobs. Humanity also collectively triumphed, as our space exploration programs broke new frontiers this year and we began to confront the increasingly real question: Should we all just move to Mars? We're just two weeks away from what is hopefully a much better 12 months, and the Engadget team took some time to commemorate our favorite gadgets and trends in tech.

In a world full of smart devices, self-driving cars and voice assistants, Shazam is the closest technology comes to actual magic. The software allows you to hold your phone up to a speaker and answer the age-old question, "what is this song and who's it by?" without the humiliation of having to ask the DJ. And in 2018 the answer, most frequently, was "Solo by Clean Bandit". The song, which features US star Demi Lovato, was tagged 9.1 million times. British artists performed five of Shazam's top 10 songs, with Calvin Harris, Dua Lipa and newcomer Tom Walker all making the chart.

Makes me feel sad for the rest. Actually, that's a movie ("The Spy that Loved Me") that Netflix recommends for me since I'm a James Bond junkie and Netflix knows that. In fact, Netflix knows a lot about me as it knows a lot about all of its viewers, which is one reason why Netflix is a Wall Street darling and has rewarded its stockholders very well over the past couple of years (see Figure 1). But Netflix isn't doing anything that other organizations cannot do. To replicate Netflix's business success starts with thinking differently about the role of data and analytics in powering the organization's business.

In his bestselling book, Up the Organization, former Avis president Robert Townsend captured the problem of automation precisely. Writing at a time when the vast paper systems of corporate America were being transferred to computers, he warned that it was important first to make sure that a company's paper systems are actually effective and accurate. "Otherwise," he quipped, "your new computer will just speed up the mess." Today, we are faced with a new wave of optimism about the prospects of what is called artificial intelligence (AI). It is important to parse these words carefully for they will tell you why artificial intelligence as it is currently conceived will very likely "just speed up the mess."

We consider the problem of matrix completion with side information on an $n\times m$ matrix. We formulate the problem exactly as a sparse regression problem of selecting features and show that it can be reformulated as a binary convex optimization problem. We design OptComplete, based on a novel concept of stochastic cutting planes to enable efficient scaling of the algorithm up to matrices of sizes $n = 10^6$ and $m = 10^5$. We report experiments on both synthetic and real-world datasets that show that OptComplete outperforms current state-of-the-art methods both in terms of accuracy and scalability, while providing insight on the factors that affect the ratings.

As deep learning advances, algorithms of music composition increase in performance. However, most of the successful models are designed for specific musical structures. Here, we present BachProp, an algorithmic composer that can generate music scores in many styles given sufficient training data. To adapt BachProp to a broad range of musical styles, we propose a novel representation of music and train a deep network to predict the note transition probabilities of a given music corpus. In this paper, new music scores generated by BachProp are compared with the original corpora as well as with different network architectures and other related models. We show that BachProp captures important features of the original datasets better than other models and invite the reader to a qualitative comparison on a large collection of generated songs.

We propose a novel data augmentation method for labeled sentences called conditional BERT contextual augmentation. Data augmentation methods are often applied to prevent overfitting and improve generalization of deep neural network models. Recently proposed contextual augmentation augments labeled sentences by randomly replacing words with more varied substitutions predicted by language model. BERT demonstrates that a deep bidirectional language model is more powerful than either an unidirectional language model or the shallow concatenation of a forward and backward model. We retrofit BERT to conditional BERT by introducing a new conditional masked language model\footnote{The term "conditional masked language model" appeared once in original BERT paper, which indicates context-conditional, is equivalent to term "masked language model". In our paper, "conditional masked language model" indicates we apply extra label-conditional constraint to the "masked language model".} task. The well trained conditional BERT can be applied to enhance contextual augmentation. Experiments on six various different text classification tasks show that our method can be easily applied to both convolutional or recurrent neural networks classifier to obtain obvious improvement.

Panoply, the smart cloud data warehouse built for business intelligence, is excited to announce its status as a finalist in Microsoft and Calcalist's Artificial Intelligence and Big Data Startup Competition. In a stiff competition, Microsoft will host representatives from Panoply and the other finalists at its Seattle headquarters. The first and second place winners will be invited by Calcalist to participate in the newspaper's Berlin conference in 2019, where the companies will meet with potential partners and investors. Panoply's CEO and co-Founder Yaniv Leven said, "We're proud to be announced as a finalist in the Artificial Intelligence and Big Data Startup competition. So far, we've been amazed at the quality of the companies participating and we're psyched to be pitching at the finals. We're a team of engineers and fighters, we set lofty goals and conquer massive challenges to prove how bad our innovative desire is."

If you thought you were going to make it out of 2018 without a couple more data slip-ups, think again! Monday, Google revealed that a bug in its somehow still alive Google social network exposed the data of 52.5 million users. That's orders of magnitude bigger than the 500,000 users that were impacted by a previous Google exposure. And on Friday, Facebook announced that it had exposed photos of up to 6.8 million users for nearly two weeks in September. The timing on Facebook's disclosure was auspicious! Not only had it just opened a one-day "pop-up" in New York City to tout its focus on user privacy, it had also announced its biggest yet bug bounty payout.