Researchers from Facebook and MIT Labs have proposed a new methodology that uses machine learning and satellite imagery to generate street addresses in areas of the world where individual buildings don't have a unique address. The methodology divides the street addressing into two processes. The first process is segmentation. During segmentation, road pixels are identified using deep learning from 0.5 meter resolution satellite images. The second part of segmentation involves developing the road network from these identified pixels.

On the other hand we read now about online technologies, combined with big data and artificial intelligence, becoming "the autocrat's new toolkit". The same technologies we enjoy - voice recognition, intelligent applications, wayfinding, habit-based recommendations - can now be used to keep track of our location, our communications and our personal lives. Every advance seems to have a potential downside. The very same NHS-DeepMind collaboration which has been improving diagnoses, has drawn privacy concerns about DeepMind's sharing of UK data with its US-based parent company Google. Citizens, businesses and governments need to talk about these issues, or the technology will run ahead of what societies want or will tolerate.

Every day, our kids are swept through the world by algorithms. YouTube algorithms decide what videos they watch, GPS algorithms map what route they take to school, Spotify algorithms select what songs they hear, and personal assistants like Siri and Alexa advise them -- all of it driven by artificial intelligence. Kids (and adults!) leave these passive engagements with AI without any physical product -- just an endless stream of passive consumption. Instead of getting carried away by these digital currents, teachers, parents, and caregivers should show kids how to experiment with powerful tools like machine learning and neural networks. We must raise kids who are capable of working alongside artificial intelligence in the workplace.

The widespread use of neural networks across different scientific domains often involves constraining them to satisfy certain symmetries, conservation laws, or other domain knowledge. Such constraints are often imposed as soft penalties during model training and effectively act as domain-specific regularizers of the empirical risk loss. Physics-informed neural networks is an example of this philosophy in which the outputs of deep neural networks are constrained to approximately satisfy a given set of partial differential equations. In this work we review recent advances in scientific machine learning with a specific focus on the effectiveness of physics-informed neural networks in predicting outcomes of physical systems and discovering hidden physics from noisy data. We will also identify and analyze a fundamental mode of failure of such approaches that is related to numerical stiffness leading to unbalanced back-propagated gradients during model training. To address this limitation we present a learning rate annealing algorithm that utilizes gradient statistics during model training to balance the interplay between different terms in composite loss functions. We also propose a novel neural network architecture that is more resilient to such gradient pathologies. Taken together, our developments provide new insights into the training of constrained neural networks and consistently improve the predictive accuracy of physics-informed neural networks by a factor of 50-100x across a range of problems in computational physics. All code and data accompanying this manuscript are publicly available at \url{https://github.com/PredictiveIntelligenceLab/GradientPathologiesPINNs}.

Like humans, deep networks learn better when samples are organized and introduced in a meaningful order or curriculum (Weinshall et al., 2018). While con-ventional approaches to curriculum learning emphasize the difficulty of samples as the core incremental strategy, it forces networks to learn from small subsets of data while introducing pre-computation overheads. In this work, we propose Learning with Incremental Labels and Adaptive Compensation(LILAC), which takes a novel approach to curriculum learning. LILAC emphasizes incrementally learning labels instead of incrementally learning difficult samples. It works in two distinct phases: first, in the incremental label introduction phase, we recursively reveal ground-truth labels in small installments while using a fake label for the remaining data. In the adaptive compensation phase, we compensate for failed predictions by adaptively altering the target vector to a smoother distribution. We evaluate LILAC against the closest comparable methods in batch and curriculum learning and label smoothing, across three standard image benchmarks, CIFAR-10, CIFAR-100, and STL-10. We show that our method outperforms batch learning with higher mean recognition accuracy as well as lower standard deviation in performance consistently across all benchmarks. We further extend LILAC to show the highest performance on CIFAR-10 for methods using simple data augmentation while exhibiting label-order invariance among other properties.

In the context of science, the well-known adage "a picture is worth a thousand words" might well be "a model is worth a thousand datasets." Scientific models, such as Newtonian physics or biological gene regulatory networks, are human-driven simplifications of complex phenomena that serve as surrogates for the countless experiments that validated the models. Recently, machine learning has been able to overcome the inaccuracies of approximate modeling by directly learning the entire set of nonlinear interactions from data. However, without any predetermined structure from the scientific basis behind the problem, machine learning approaches are flexible but data-expensive, requiring large databases of homogeneous labeled training data. A central challenge is reconciling data that is at odds with simplified models without requiring "big data". In this work we develop a new methodology, universal differential equations (UDEs), which augments scientific models with machine-learnable structures for scientifically-based learning. We show how UDEs can be utilized to discover previously unknown governing equations, accurately extrapolate beyond the original data, and accelerate model simulation, all in a time and data-efficient manner. This advance is coupled with open-source software that allows for training UDEs which incorporate physical constraints, delayed interactions, implicitly-defined events, and intrinsic stochasticity in the model. Our examples show how a diverse set of computationally-difficult modeling issues across scientific disciplines, from automatically discovering biological mechanisms to accelerating climate simulations by 15,000x, can be handled by training UDEs.

January 15, 2020 Abstract This paper presents a time series forecasting framework which combines standard forecasting methods and a machine learning model. The inputs to the machine learning model are not lagged values or regular time series features, but instead forecasts produced by standard methods. The machine learning model can be either a convolutional neural network model or a recurrent neural network model. The intuition behind this approach is that forecasts of a time series are themselves good features characterizing the series, especially when the modelling purpose is forecasting. It can also be viewed as a weighted ensemble method. Tested on the M4 competition dataset, this approach outperforms all submissions for quarterly series, and is more accurate than all but the winning algorithm for monthly series. 1 Introduction The competitiveness of neural network (NN) models and other machine learning (ML) models for time series forecasting compared to statistical models has long been questioned by practitioners [1] [2]. Although in the field of time series forecasting, there is a plethora of literature presenting complex novel models, in practice the performance of ML models is often below expectation [3].

Typical Magnetic Resonance Imaging (MRI) scan may take 20 to 60 minutes. Reducing MRI scan time is beneficial for both patient experience and cost considerations. Accelerated MRI scan may be achieved by acquiring less amount of k-space data (down-sampling in the k-space). However, this leads to lower resolution and aliasing artifacts for the reconstructed images. There are many existing approaches for attempting to reconstruct high-quality images from down-sampled k-space data, with varying complexity and performance. In recent years, deep-learning approaches have been proposed for this task, and promising results have been reported. Still, the problem remains challenging especially because of the high fidelity requirement in most medical applications employing reconstructed MRI images. In this work, we propose a deep-learning approach, aiming at reconstructing high-quality images from accelerated MRI acquisition. Specifically, we use Convolutional Neural Network (CNN) to learn the differences between the aliased images and the original images, employing a U-Net-like architecture. Further, a micro-architecture termed Residual Dense Block (RDB) is introduced for learning a better feature representation than the plain U-Net. Considering the peculiarity of the down-sampled k-space data, we introduce a new term to the loss function in learning, which effectively employs the given k-space data during training to provide additional regularization on the update of the network weights. To evaluate the proposed approach, we compare it with other state-of-the-art methods. In both visual inspection and evaluation using standard metrics, the proposed approach is able to deliver improved performance, demonstrating its potential for providing an effective solution.

Large Transformer models routinely achieve state-of-the-art results on a number of tasks but training these models can be prohibitively costly, especially on long sequences. We introduce two techniques to improve the efficiency of Transformers. Furthermore, we use reversible residual layers instead of the standard residuals, which allows storing activations only once in the training process instead of N times, where N is the number of layers. The resulting model, the Reformer, performs on par with Transformer models while being much more memory-efficient and much faster on long sequences. The Transformer architecture (V aswani et al., 2017) is widely used in natural language processing and yields state-of-the-art results on a number of tasks. To obtain these results, researchers have resorted to training ever larger Transformer models. The number of parameters exceeds 0.5B per layer in the largest configuration reported in (Shazeer et al., 2018) while the number of layers goes up to 64 in (Al-Rfou et al., 2018). Transformer models are also used on increasingly long sequences.

How does a 110-layer ResNet learn a high-complexity classifier using relatively few training examples and short training time? We present a theory towards explaining this in terms of $\textit{hierarchical learning}$. We refer hierarchical learning as the learner learns to represent a complicated target function by decomposing it into a sequence of simpler functions to reduce sample and time complexity. This paper formally analyzes how multi-layer neural networks can perform such hierarchical learning efficiently and automatically simply by applying stochastic gradient descent (SGD). On the conceptual side, we present, to the best of our knowledge, the FIRST theory result indicating how very deep neural networks can still be sample and time efficient on certain hierarchical learning tasks, when NO KNOWN non-hierarchical algorithms (such as kernel method, linear regression over feature mappings, tensor decomposition, sparse coding) are efficient. We establish a new principle called "backward feature correction", which we believe is the key to understand the hierarchical learning in multi-layer neural networks. On the technical side, we show for regression and even for binary classification, for every input dimension $d > 0$, there is a concept class consisting of degree $\omega(1)$ multi-variate polynomials so that, using $\omega(1)$-layer neural networks as learners, SGD can learn any target function from this class in $\mathsf{poly}(d)$ time using $\mathsf{poly}(d)$ samples to any $\frac{1}{\mathsf{poly}(d)}$ error, through learning to represent it as a composition of $\omega(1)$ layers of quadratic functions. In contrast, we present lower bounds stating that several non-hierarchical learners, including any kernel methods, neural tangent kernels, must suffer from $d^{\omega(1)}$ sample or time complexity to learn functions in this concept class even to any $d^{-0.01}$ error.