A major challenge in machine learning is the computational expense of training these models. Model training can be viewed as a form of optimization used to fit a machine learning model to a set of data, which can take up significant amount of time on classical computers. Adiabatic quantum computers have been shown to excel at solving optimization problems, and therefore, we believe, present a promising alternative to improve machine learning training times. In this paper, we present an adiabatic quantum computing approach for training a linear regression model. In order to do this, we formulate the regression problem as a quadratic unconstrained binary optimization (QUBO) problem. We analyze our quantum approach theoretically, test it on the D-Wave 2000Q adiabatic quantum computer and compare its performance to a classical approach that uses the Scikit-learn library in Python. Our analysis shows that the quantum approach attains up to 2.8x speedup over the classical approach on larger datasets, and performs at par with the classical approach on the regression error metric.

Recent work has attempted to interpret residual networks (ResNets) as one step of a forward Euler discretization of an ordinary differential equation, focusing mainly on syntactic algebraic similarities between the two systems. Discrete dynamical integrators of continuous dynamical systems, however, have a much richer structure. We first show that ResNets fail to be meaningful dynamical integrators in this richer sense. We then demonstrate that neural network models can learn to represent continuous dynamical systems, with this richer structure and properties, by embedding them into higher-order numerical integration schemes, such as the Runge Kutta schemes. Based on these insights, we introduce ContinuousNet as a continuous-in-depth generalization of ResNet architectures. ContinuousNets exhibit an invariance to the particular computational graph manifestation. That is, the continuous-in-depth model can be evaluated with different discrete time step sizes, which changes the number of layers, and different numerical integration schemes, which changes the graph connectivity. We show that this can be used to develop an incremental-in-depth training scheme that improves model quality, while significantly decreasing training time. We also show that, once trained, the number of units in the computational graph can even be decreased, for faster inference with little-to-no accuracy drop.

We present a Bayesian approach to identify optimal transformations that map model input points to low dimensional latent variables. The "projection" mapping consists of an orthonormal matrix that is considered a priori unknown and needs to be inferred jointly with the GP parameters, conditioned on the available training data. The proposed Bayesian inference scheme relies on a two-step iterative algorithm that samples from the marginal posteriors of the GP parameters and the projection matrix respectively, both using Markov Chain Monte Carlo (MCMC) sampling. In order to take into account the orthogonality constraints imposed on the orthonormal projection matrix, a Geodesic Monte Carlo sampling algorithm is employed, that is suitable for exploiting probability measures on manifolds. We extend the proposed framework to multi-fidelity models using GPs including the scenarios of training multiple outputs together. We validate our framework on three synthetic problems with a known lower-dimensional subspace. The benefits of our proposed framework, are illustrated on the computationally challenging three-dimensional aerodynamic optimization of a last-stage blade for an industrial gas turbine, where we study the effect of an 85-dimensional airfoil shape parameterization on two output quantities of interest, specifically on the aerodynamic efficiency and the degree of reaction.

Training machine learning models on classical computers is usually a time and compute intensive process. With Moore's law coming to an end and ever increasing demand for large-scale data analysis using machine learning, we must leverage non-conventional computing paradigms like quantum computing to train machine learning models efficiently. Adiabatic quantum computers like the D-Wave 2000Q can approximately solve NP-hard optimization problems, such as the quadratic unconstrained binary optimization (QUBO), faster than classical computers. Since many machine learning problems are also NP-hard, we believe adiabatic quantum computers might be instrumental in training machine learning models efficiently in the post Moore's law era. In order to solve a problem on adiabatic quantum computers, it must be formulated as a QUBO problem, which is a challenging task in itself. In this paper, we formulate the training problems of three machine learning models---linear regression, support vector machine (SVM) and equal-sized k-means clustering---as QUBO problems so that they can be trained on adiabatic quantum computers efficiently. We also analyze the time and space complexities of our formulations and compare them to the state-of-the-art classical algorithms for training these machine learning models. We show that the time and space complexities of our formulations are better (in the case of SVM and equal-sized k-means clustering) or equivalent (in case of linear regression) to their classical counterparts.

CAMBRIDGE – COVID-19 has become a severe stress test for countries around the world. From supply-chain management and health-care capacity to regulatory reform and economic stimulus, the pandemic has mercilessly punished governments that did not – or could not – adapt quickly. From Latin America's lost decade in the 1980s to the more recent Greek crisis, there are plenty of painful reminders of what happens when countries cannot service their debts. A global debt crisis today would likely push millions of people into unemployment and fuel instability and violence around the world. The virus has also pulled back the curtain on one of this century's most important contests: the rivalry between the United States and China for supremacy in artificial intelligence (AI).

The world needs robots that make life better, not just ones that put people out of work. But business attitudes, government policy, and scientific priorities are geared toward replacing workers rather than complementing and enhancing their skills. That's the bottom line of a report by a task force at MIT that was released today. "It's super easy to make a business case for reducing head count. You can always light up a boardroom" by promising to replace people with robots, says David Autor, an MIT economist and co-chair of the task force, who gave an interview about the report.

Today, Buolamwini is galvanizing a growing movement to expose the social consequences of artificial intelligence. Through her nearly four-year-old nonprofit, the Algorithmic Justice League (AJL), she has testified before lawmakers at the federal, state, and local levels about the dangers of using facial recognition technologies with no oversight of how they're created or deployed. Since George Floyd's death, she has called for a complete halt to police use of face surveillance, and is providing activists with resources and tools to demand regulation. Many companies, such as Clearview AI, are still selling facial analysis to police and government agencies. And many police departments are using facial recognition technologies to identify, in the words of the New York Police Department, individuals that have committed, are committing, or are about to commit crimes.

One of the challenges with modern machine learning systems is that they are very heavily dependent on large quantities of data to make them work well. This is especially the case with deep neural nets, where lots of layers means lots of neural connections which requires large amounts of data and training to get to the point where the system can provide results at acceptable levels of accuracy and precision. Indeed, the ultimate implementation of this massive data, massive network vision is the currently much-vaunted Open AI GPT-3, which is so large that it can predict and generate almost any text with surprising magical wizardry. However, in many ways, GPT-3 is still a big data magic trick. Indeed, Professor Luis Perez-Breva makes this exact point when he says that what we call machine learning isn't really learning at all.

As artificial intelligence and machine learning-based systems become more ubiquitous in decision-making, should we expect our confidence in the outcomes to remain like that of its human collaborators? When humans make decisions, we're able to rationalize the outcomes through inquiry and conversation around how expert judgment, experience and use of available information led to the decision. To borrow the words of former Secretary of Defense Ash Carter when speaking at a 2019 SXSW panel about post-analysis of an AI-enabled decision, "'the machine did it' won't fly." As we evolve human and machine collaboration, establishing trust, transparency and accountability at the onset of decision support system and algorithm design is paramount. Without it, people may be hesitant to trust AI recommendations because of a lack of transparency into how the machine reached its outcome.

Every time Congress holds a hearing about Silicon Valley companies, people mock the legislators for being out of their depth. Last week's effort by the antitrust subcommittee of the House Judiciary Committee was no exception. "The technological ignorance demonstrated by our elected officials ... was truly stunning," Shelly Palmer, CEO at the Palmer Group, a tech strategy advisory group, told USA Today. "People who are this clueless about the economic forces shaping our world should not be tasked with leading us into the age of AI," he said. "The data elite are playing a different game with a different set of rules. Apparently, Congress can't even find the ballpark."