If you haven't heard about the artificial intelligence (AI) machine-learning (ML) craze that uses deep neural networks (DNN) and deep learning (DL) to tackle everything from voice recognition to making self-driving cars a reality, then you probably haven't heard about Google's new Tensor Processing Unit (TPU), Intel's Lake Crest, or Knupath's Hermosa. These are just a few of the vendors looking to deliver platforms targeting neural networks. The TPU contains a large 8-bit matrix multiply unit (Figure 1). It essentially optimizes the number-crunching required by DNN; large floating-point number-crunchers need not apply. The TPU is actually a coprocessor managed by a conventional host CPU via the TPU's PCI Express interface.

When billionaire entrepreneurs, investors and technologists talk, people listen. Vinod Khosla is all of those things--and he's predicting the imminent demise of radiology at the hands of AI algorithms. "The role of the radiologist will be obsolete in five years," Khosla tells CNBC reporter Christina Farr. Garry Choy, MD, MBA, says that, far from replacing rads, AI technology will free them up to do more high-order, complex thinking as well as spend more time with patients. And Simon Rascovsky, MD, MS, tells Farr the specialty has heard it all before and still doesn't have "clinically proven deep learning based applications outside of pilots and marketing hype."

Alphabet's artificial intelligence (AI) division DeepMind continues to enhance its healthcare credentials with the arrival of a Bitcoin-like auditing process to protect the personal data of patients. The London-based firm is calling its data verification process the "Verifiable Data Audit". The idea is that it will give patients peace of mind with a "mathematical assurance about what is happening with each individual piece of personal data." It comes amid growing awareness of the value of personal data, and a loss in trust in organisation's abilities to protect their information. The NHS for example is notorious over its shoddy handling of patient data, with numerous incidents reported over the past few years.

Just over a year ago, an artificial intelligence beat the fourth-ranked player in the world at Go, a complex and ancient game that is said to be a better test of the unique capabilities of the human brain than chess. Now, that same artificial intelligence, called AlphaGo, is preparing for its next public demonstration: a summit in China in May where it will collaborate with human players to come up with strategies, and then face off against the top-ranked player in a series of three matches. If it wins, it will have shown that its underlying algorithms are ready for something more than a game. AlphaGo is the product of DeepMind, a London-based division of Google parent company Alphabet. DeepMind was founded in 2010 around the goal of researching artificial intelligence in order to understand the nature of intelligence and harness the full power of computerized brainpower for humanity. Its experiments with Go -- a game thought to be years away from being conquered by AI before last year -- are designed to bring us closer to designing a computer with human-like understanding that can solve problems like a human mind can.

The choice depend on your current and future needs.

This week's top Machine Learning stories, including computational models of drug effectiveness, stem cells, sentiment analysis, and more! Machine Learning is one of the most exciting fields in the world. Every week we discover something new, something amazing, something revolutionary. It's incredible, but it can also be overwhelming. That's why we created This Week in Machine Learning!

One of the biggest challenges facing the development of AI is that it requires a huge amount of human input, both in terms of the involvement of people when it comes to identifying and inputting data up front, and in terms of the nature of data sets required to even make training AI systems possible to begin with. Google AI research Ian Goodfellow, who recently headed back to Google Brain after a stint at the Elon Musk-backed OpenAI, hopes to address both those issues through an approach to AI that involves pitting one neural network against another. The concept isn't new: Facebook published a paper co-authored by its head of AI research Yann LeCunn and AI engineer Soumith Chintala last June, in which they describe using generative adversarial networks (GANs) to eventually enable unsupervised learning, aka machine learning that takes place without any human involvement. Goodfellow pioneered this idea, however, proving its basic viability after a heated (and boozy) debate with some University of Montreal academic colleagues, as Wired reports. In essence, the nature of the system includes two opposing neural networks that inform one another through their opposition: the first tries to create something synthetic, for instance a realistic image of a dog, and the other criticizes its attempts, trying to spot the fakes and pointing out where the first system has failed or fallen down.

In this paper, we propose a novel generative model named Stacked Generative Adversarial Networks (SGAN), which is trained to invert the hierarchical representations of a bottom-up discriminative network. Our model consists of a top-down stack of GANs, each learned to generate lower-level representations conditioned on higher-level representations. A representation discriminator is introduced at each feature hierarchy to encourage the representation manifold of the generator to align with that of the bottom-up discriminative network, leveraging the powerful discriminative representations to guide the generative model. In addition, we introduce a conditional loss that encourages the use of conditional information from the layer above, and a novel entropy loss that maximizes a variational lower bound on the conditional entropy of generator outputs. We first train each stack independently, and then train the whole model end-to-end. Unlike the original GAN that uses a single noise vector to represent all the variations, our SGAN decomposes variations into multiple levels and gradually resolves uncertainties in the top-down generative process. Based on visual inspection, Inception scores and visual Turing test, we demonstrate that SGAN is able to generate images of much higher quality than GANs without stacking.

This paper presents sampling-based speech parameter generation using moment-matching networks for Deep Neural Network (DNN)-based speech synthesis. Although people never produce exactly the same speech even if we try to express the same linguistic and para-linguistic information, typical statistical speech synthesis produces completely the same speech, i.e., there is no inter-utterance variation in synthetic speech. To give synthetic speech natural inter-utterance variation, this paper builds DNN acoustic models that make it possible to randomly sample speech parameters. The DNNs are trained so that they make the moments of generated speech parameters close to those of natural speech parameters. Since the variation of speech parameters is compressed into a low-dimensional simple prior noise vector, our algorithm has lower computation cost than direct sampling of speech parameters. As the first step towards generating synthetic speech that has natural inter-utterance variation, this paper investigates whether or not the proposed sampling-based generation deteriorates synthetic speech quality. In evaluation, we compare speech quality of conventional maximum likelihood-based generation and proposed sampling-based generation. The result demonstrates the proposed generation causes no degradation in speech quality.

Learning based methods have shown very promising results for the task of depth estimation in single images. However, most existing approaches treat depth prediction as a supervised regression problem and as a result, require vast quantities of corresponding ground truth depth data for training. Just recording quality depth data in a range of environments is a challenging problem. In this paper, we innovate beyond existing approaches, replacing the use of explicit depth data during training with easier-to-obtain binocular stereo footage. We propose a novel training objective that enables our convolutional neural network to learn to perform single image depth estimation, despite the absence of ground truth depth data. Exploiting epipolar geometry constraints, we generate disparity images by training our network with an image reconstruction loss. We show that solving for image reconstruction alone results in poor quality depth images. To overcome this problem, we propose a novel training loss that enforces consistency between the disparities produced relative to both the left and right images, leading to improved performance and robustness compared to existing approaches. Our method produces state of the art results for monocular depth estimation on the KITTI driving dataset, even outperforming supervised methods that have been trained with ground truth depth.