When someone commits suicide, their family and friends can be left with the heartbreaking and answerless question of what they could have done differently. Colin Walsh, data scientist at Vanderbilt University Medical Center, hopes his work in predicting suicide risk will give people the opportunity to ask "what can I do?" while there's still a chance to intervene. Walsh and his colleagues have created machine-learning algorithms that predict, with unnerving accuracy, the likelihood that a patient will attempt suicide. In trials, results have been 80-90% accurate when predicting whether someone will attempt suicide within the next two years, and 92% accurate in predicting whether someone will attempt suicide within the next week. The prediction is based on data that's widely available from all hospital admissions, including age, gender, zip codes, medications, and prior diagnoses.

Scientists, data scientists that is, from the University of Adelaide in Australia have announced that they have managed to build an Artificial Intelligence (AI) that can predict when people are going to die, and it's 70 percent accurate, but unlike the AI's I've talked about before that can predict how long people who have had heart attacks have left to live, more accurately than human doctors, this one is different – it can predict when you're going to die irrespective of the state of your current health because it uses deep learning to analyse a range of different scans, such as CT scans, to search for the signs, and assess the severity of, heart disease, cancer, and other diseases. For example, look at it this way – if I asked you how long you thought you were going to live, and you exercised every day, had the perfect diet and had no history of hereditary diseases in your family's lineage then it's likely you'd be able to tell me, with some degree of confidence, that your chances of living beyond eighty years old were good. On the other hand though, if you never exercised, had an awful diet of alcohol, lard and sugar, and your family had a history of hereditary diseases then you might tell me that the chances of you living beyond eighty could be slim. See, in one fell swoop you've assessed the state of your overall health, roughly assessed the risk factors in your head and calculated the rough odds of how long you think you have left to live. And that's what this AI is being trained to do, except for the fact that rather than relying on gut instinct it's analysing real time scans and correlating the patterns it's seeing against a huge dataset of patient information.

Thinking about how and when you'll die might be morbid, but it has creeped into everyone's mind at some point. Online tools like The Death Clock provide a very unscientific, and entertaining, prediction of your demise, but researchers have figured out a way to estimate a person's lifespan with 69 percent accuracy. In a very small study of 48 participants, all of whom were at least 60 years old, scientists from the University of Adelaide in Australia analyzed photos of people's organs using artificial intelligence. They were able to predict who would die within five years with 69 percent accuracy, which is roughly the same as an oncologist's. Using deep learning, which involves inputting data into a computer system to help it make decisions, the researchers used radiological images because they provide undetectable clues, according to study co-author and epidemiologist Dr. Lyle Palmer, Ph.D, in a story on ResearchGate.

Earnix, a provider of analytics solutions for the financial services industry, today announced the results of a global survey of insurance executives, which shows wide adoption of Machine Learning across the globe, and the expectation that ML will bring "significant" change to the industry over the next three to five years. Over half (54%) of the almost 200 insurance executives surveyed said that their organization was using Machine Learning for predictive analytical modelling. Of those deploying the technology, 70% said they were using it for risk modelling; followed by demand models (45%) and fraud detection (36%). Although nascent, most companies using Machine Learning have realized measurable benefits. Over half of the respondents (57%) said that Machine Learning has made their analytical models far more accurate, which has led to better risk assessments, and ultimately better decisions.

Neural networks and rational functions efficiently approximate each other. In more detail, it is shown here that for any ReLU network, there exists a rational function of degree $O(\text{polylog}(1/\epsilon))$ which is $\epsilon$-close, and similarly for any rational function there exists a ReLU network of size $O(\text{polylog}(1/\epsilon))$ which is $\epsilon$-close. By contrast, polynomials need degree $\Omega(\text{poly}(1/\epsilon))$ to approximate even a single ReLU. When converting a ReLU network to a rational function as above, the hidden constants depend exponentially on the number of layers, which is shown to be tight; in other words, a compositional representation can be beneficial even for rational functions.

Deep learning models can take weeks to train on a single GPU-equipped machine, necessitating scaling out DL training to a GPU-cluster. However, current distributed DL implementations can scale poorly due to substantial parameter synchronization over the network, because the high throughput of GPUs allows more data batches to be processed per unit time than CPUs, leading to more frequent network synchronization. We present Poseidon, an efficient communication architecture for distributed DL on GPUs. Poseidon exploits the layered model structures in DL programs to overlap communication and computation, reducing bursty network communication. Moreover, Poseidon uses a hybrid communication scheme that optimizes the number of bytes required to synchronize each layer, according to layer properties and the number of machines. We show that Poseidon is applicable to different DL frameworks by plugging Poseidon into Caffe and TensorFlow. We show that Poseidon enables Caffe and TensorFlow to achieve 15.5x speed-up on 16 single-GPU machines, even with limited bandwidth (10GbE) and the challenging VGG19-22K network for image classification. Moreover, Poseidon-enabled TensorFlow achieves 31.5x speed-up with 32 single-GPU machines on Inception-V3, a 50% improvement over the open-source TensorFlow (20x speed-up).

We consider the problem of training generative models with a Generative Adversarial Network (GAN). Although GANs can accurately model complex distributions, they are known to be difficult to train due to instabilities caused by a difficult minimax optimization problem. In this paper, we view the problem of training GANs as finding a mixed strategy in a zero-sum game. Building on ideas from online learning we propose a novel training method named Chekhov GAN 1 . On the theory side, we show that our method provably converges to an equilibrium for semi-shallow GAN architectures, i.e. architectures where the discriminator is a one layer network and the generator is arbitrary. On the practical side, we develop an efficient heuristic guided by our theoretical results, which we apply to commonly used deep GAN architectures. On several real world tasks our approach exhibits improved stability and performance compared to standard GAN training.

We consider maximum likelihood estimation for Gaussian Mixture Models (Gmms). This task is almost invariably solved (in theory and practice) via the Expectation Maximization (EM) algorithm. EM owes its success to various factors, of which is its ability to fulfill positive definiteness constraints in closed form is of key importance. We propose an alternative to EM by appealing to the rich Riemannian geometry of positive definite matrices, using which we cast Gmm parameter estimation as a Riemannian optimization problem. Surprisingly, such an out-of-the-box Riemannian formulation completely fails and proves much inferior to EM. This motivates us to take a closer look at the problem geometry, and derive a better formulation that is much more amenable to Riemannian optimization. We then develop (Riemannian) batch and stochastic gradient algorithms that outperform EM, often substantially. We provide a non-asymptotic convergence analysis for our stochastic method, which is also the first (to our knowledge) such global analysis for Riemannian stochastic gradient. Numerous empirical results are included to demonstrate the effectiveness of our methods.

I consider unsupervised extensions of the fast stepwise linear regression algorithm \cite{efroymson1960multiple}. These extensions allow one to efficiently identify highly-representative feature variable subsets within a given set of jointly distributed variables. This in turn allows for the efficient dimensional reduction of large data sets via the removal of redundant features. Fast search is effected here through the avoidance of repeat computations across trial fits, allowing for a full representative-importance ranking of a set of feature variables to be carried out in $O(n^2 m)$ time, where $n$ is the number of variables and $m$ is the number of data samples available. This runtime complexity matches that needed to carry out a single regression and is $O(n^2)$ faster than that of naive implementations. I present pseudocode suitable for efficient forward, reverse, and forward-reverse unsupervised feature selection. To illustrate the algorithm's application, I apply it to the problem of identifying representative stocks within a given financial market index -- a challenge relevant to the design of Exchange Traded Funds (ETFs). I also characterize the growth of numerical error with iteration step in these algorithms, and finally demonstrate and rationalize the observation that the forward and reverse algorithms return exactly inverted feature orderings in the weakly-correlated feature set regime.

We address the statistical and optimization impacts of using classical sketch versus Hessian sketch to solve approximately the Matrix Ridge Regression (MRR) problem. Prior research has considered the effects of classical sketch on least squares regression (LSR), a strictly simpler problem. We establish that classical sketch has a similar effect upon the optimization properties of MRR as it does on those of LSR---namely, it recovers nearly optimal solutions. In contrast, Hessian sketch does not have this guarantee, instead, the approximation error is governed by a subtle interplay between the "mass" in the responses and the optimal objective value. For both types of approximations, the regularization in the sketched MRR problem gives it significantly different statistical properties from the sketched LSR problem. In particular, there is a bias-variance trade-off in sketched MRR that is not present in sketched LSR. We provide upper and lower bounds on the biases and variances of sketched MRR, these establish that the variance is significantly increased when classical sketches are used, while the bias is significantly increased when using Hessian sketches. Empirically, sketched MRR solutions can have risks that are higher by an order-of-magnitude than those of the optimal MRR solutions. We establish theoretically and empirically that model averaging greatly decreases this gap. Thus, in the distributed setting, sketching combined with model averaging is a powerful technique that quickly obtains near-optimal solutions to the MRR problem while greatly mitigating the statistical risks incurred by sketching.