"Deep learning" is the new buzzword in the field of artificial intelligence. As Natalie Wolchover reported in a recent Quanta Magazine article, "'deep neural networks' have learned to converse, drive cars, beat video games and Go champions, dream, paint pictures and help make scientific discoveries." With such successes, one would expect deep learning to be a revolutionary new technique. But one would be quite wrong. The basis of deep learning stretches back more than half a century to the dawn of AI and the creation of both artificial neural networks having layers of connected neuronlike units and the "back propagation algorithm" -- a technique of applying error corrections to the strengths of the connections between neurons on different layers.

Arterys is working on applying artificial intelligence to create algorithms that will impact the lives of millions of people. We are looking for passionate and talented deep learning experts to advance the field of medicine. Much of radiology consists of performing tedious tasks that would greatly benefit from automation, such as segmenting anatomical regions, characterizing lesions and writing reports. Additionally, non-radiologist clinicians who make treatment decisions often do not have the ability to synthesize the myriad components of available clinical information to make the best possible treatment decisions for their patients. Arterys' goal is to use the latest machine learning technology to solve these problems, help clinicians work more efficiently, and make a dramatic impact on patient outcomes.

DeepMind, the UK's star artificial intelligence company owned by Google, has tripled the amount of money it spends on top talent. Spending on administration largely driven by its wage bill jumped to £164m in 2016, up from £54m a year earlier, according to its newly published annual accounts, as it splashed the cash on attracting and retaining experts in the highly competitive field. "We're really proud that some of the world's most exciting AI research and real-world application is taking place right here in London," said a spokesperson for DeepMind. "We intend to keep investing in our scientific mission, and to work with the world's brightest minds to tackle society's most complex problems." Read more: 5 things we learned about DeepMind's Demis Hassabis on Desert Island Discs The London-based tech firm which was snapped up by Google for £400m in 2014, brought in revenue for the first time last year of £40m from its research and development into AI, the figures also reveal.

Google's artificial intelligence research arm DeepMind has launched a unit focused on ethics and society. The group will conduct and fund research that covers the humanities and social sciences and run public discussion events, DeepMind announced earlier this week. The unit has already released five'core principals' to guide future AI research: that technologies be developed in ways that serve the global social and environmental good; that research be'rigorous and evidence-based' as well as'transparent and open' (including with funding arrangements); that work includes a diversity of voices; and that public opinion will feature in all developments. "This new unit will help us explore and understand the real-world impacts of AI," the group wrote in a blog post earlier this week. "It has a dual aim: to help technologists put ethics into practice, and to help society anticipate and direct the impact of AI so that it works for the benefit of all."

Every few months I enter the following query into Google: "Tensorflow sucks" or "f*** Tensorflow", hoping to find like-minded folk on the internet. Unfortunately, although Tensorflow has been around for about two years, I still cannot find a bashing of Tensorflow that leaves me fully satisfied. Although I suppose it's possible I might be asking the wrong search engine, I think there's a different force at work here: Google envy. The phenomenon known as "Google deep envy" is the following set of assumptions made by engineers across the world: Let's leave our assumptions behind us for now and give Tensorflow an honest look. When Tensorflow first came out, we were promised an end to the endless nightmare of poorly designed or poorly maintained deep learning frameworks.

This demo compares the performance of Deep Learning training of the CIFAR-10 dataset when run using HDP Yarn Containers on IBM Power Systems using only system CPU versus CPU plus NVIDIA Tesla P100 GPUs.

Denoising autoencoders (DAEs) are powerful deep learning models used for feature extraction, data generation and network pre-training. DAEs consist of an encoder and decoder which may be trained simultaneously to minimise a loss (function) between an input and the reconstruction of a corrupted version of the input. There are two common loss functions used for training autoencoders, these include the mean-squared error (MSE) and the binary cross-entropy (BCE). When training autoencoders on image data a natural choice of loss function is BCE, since pixel values may be normalised to take values in [0,1] and the decoder model may be designed to generate samples that take values in (0,1). We show theoretically that DAEs trained to minimise BCE may be used to take gradient steps in the data space towards regions of high probability under the data-generating distribution. Previously this had only been shown for DAEs trained using MSE. As a consequence of the theory, iterative application of a trained DAE moves a data sample from regions of low probability to regions of higher probability under the data-generating distribution. Firstly, we validate the theory by showing that novel data samples, consistent with the training data, may be synthesised when the initial data samples are random noise. Secondly, we motivate the theory by showing that initial data samples synthesised via other methods may be improved via iterative application of a trained DAE to those initial samples.

PET image reconstruction is challenging due to the ill-poseness of the inverse problem and limited number of detected photons. Recently deep neural networks have been widely and successfully used in computer vision tasks and attracted growing interests in medical imaging. In this work, we trained a deep residual convolutional neural network to improve PET image quality by using the existing inter-patient information. An innovative feature of the proposed method is that we embed the neural network in the iterative reconstruction framework for image representation, rather than using it as a post-processing tool. We formulate the objective function as a constraint optimization problem and solve it using the alternating direction method of multipliers (ADMM) algorithm. Both simulation data and hybrid real data are used to evaluate the proposed method. Quantification results show that our proposed iterative neural network method can outperform the neural network denoising and conventional penalized maximum likelihood methods.

In this paper we propose a function space approach to Representation Learning and the analysis of the representation layers in deep learning architectures. We show how to compute a weak-type Besov smoothness index that quantifies the geometry of the clustering in the feature space. This approach was already applied successfully to improve the performance of machine learning algorithms such as the Random Forest and tree-based Gradient Boosting. Our experiments demonstrate that in well-known and well-performing trained networks, the Besov smoothness of the training set, measured in the corresponding hidden layer feature map representation, increases from layer to layer. We also contribute to the understanding of generalization by showing how the Besov smoothness of the representations, decreases as we add more mis-labeling to the training data. We hope this approach will contribute to the de-mystification of some aspects of deep learning.

Throughout the years, research in neural networks (NN) for univariate time series forecasting has received considerable attention. Recent developments have been mainly around preprocessing techniques such as deseasonalization and detrending to supplement the NN's learning process, and novel NN architectures such as recurrent neural networks, echo state networks, generalized regression neural networks and ensemble architectures to uplift the constraints of the conventional NN architecture (Nelson et al., 1999; Zhang and Qi, 2005; Ilies et al., 2007; Rahman et al., 2016; Yan, 2012; Zimmermann et al., 2012). However, in the time series forecasting community there has also been the longstanding consensus that simple methods will often outperform more sophisticated ones. This was a conclusion of the influential M3 forecasting competition held in 1999 (Makridakis and Hibon, 2000). So, complex methods are often viewed poorly in this field, and this has been especially true for NNs and other machine learning (ML) methods. In particular, NNs did not perform well in this competition and in subsequent competitions, e.g., more recently, in the NN3 and NN5 forecasting competitions, which were held specifically for ML methods.