New generation geostationary satellites make solar reflectance observations available at a continental scale with unprecedented spatiotemporal resolution and spectral range. Generating quality land monitoring products requires correction of the effects of atmospheric scattering and absorption, which vary in time and space according to geometry and atmospheric composition. Many atmospheric radiative transfer models, including that of Multi-Angle Implementation of Atmospheric Correction (MAIAC), are too computationally complex to be run in real time, and rely on precomputed look-up tables. Additionally, uncertainty in measurements and models for remote sensing receives insufficient attention, in part due to the difficulty of obtaining sufficient ground measurements. In this paper, we present an adaptation of Bayesian Deep Learning (BDL) to emulation of the MAIAC atmospheric correction algorithm. Emulation approaches learn a statistical model as an efficient approximation of a physical model, while machine learning methods have demonstrated performance in extracting spatial features and learning complex, nonlinear mappings. We demonstrate stable surface reflectance retrieval by emulation (R2 between MAIAC and emulator SR are 0.63, 0.75, 0.86, 0.84, 0.95, and 0.91 for Blue, Green, Red, NIR, SWIR1, and SWIR2 bands, respectively), accurate cloud detection (86\%), and well-calibrated, geolocated uncertainty estimates. Our results support BDL-based emulation as an accurate and efficient (up to 6x speedup) method for approximation atmospheric correction, where built-in uncertainty estimates stand to open new opportunities for model assessment and support informed use of SR-derived quantities in multiple domains.

Memory is an important aspect of intelligence and plays a role in many deep reinforcement learning models. However, little progress has been made in understanding when specific memory systems help more than others and how well they generalize. The field also has yet to see a prevalent consistent and rigorous approach for evaluating agent performance on holdout data. In this paper, we aim to develop a comprehensive methodology to test different kinds of memory in an agent and assess how well the agent can apply what it learns in training to a holdout set that differs from the training set along dimensions that we suggest are relevant for evaluating memory-specific generalization. To that end, we first construct a diverse set of memory tasks that allow us to evaluate test-time generalization across multiple dimensions. Second, we develop and perform multiple ablations on an agent architecture that combines multiple memory systems, observe its baseline models, and investigate its performance against the task suite.

Stein's method ( Stein, 1973; 1981) is a powerful tool for statistical applications, and has had a significant impact in machine learning. Stein's lemma plays an essential role in Stein's method. Previous applications of Stein's lemma either required strong technical assumptions or were limited to Gaussian distributions with restricted covariance structures. In this work, we extend Stein's lemma to exponential-family mixture distributions including Gaussian distributions with full covariance structures. Our generalization enables us to establish a connection between Stein's lemma and the reparamterization trick to derive gradients of expectations of a large class of functions under weak assumptions. Using this connection, we can derive many new reparameterizable gradient-identities that goes beyond the reach of existing works. For example, we give gradient identities when expectation is taken with respect to Student's t-distribution, skew Gaussian, exponentially modified Gaussian, and normal inverse Gaussian.

Autonomous agents must often deal with conflicting requirements, such as completing tasks using the least amount of time/energy, learning multiple tasks, or dealing with multiple opponents. In the context of reinforcement learning~(RL), these problems are addressed by (i)~designing a reward function that simultaneously describes all requirements or (ii)~combining modular value functions that encode them individually. Though effective, these methods have critical downsides. Designing good reward functions that balance different objectives is challenging, especially as the number of objectives grows. Moreover, implicit interference between goals may lead to performance plateaus as they compete for resources, particularly when training on-policy. Similarly, selecting parameters to combine value functions is at least as hard as designing an all-encompassing reward, given that the effect of their values on the overall policy is not straightforward. The later is generally addressed by formulating the conflicting requirements as a constrained RL problem and solved using Primal-Dual methods. These algorithms are in general not guaranteed to converge to the optimal solution since the problem is not convex. This work provides theoretical support to these approaches by establishing that despite its non-convexity, this problem has zero duality gap, i.e., it can be solved exactly in the dual domain, where it becomes convex. Finally, we show this result basically holds if the policy is described by a good parametrization~(e.g., neural networks) and we connect this result with primal-dual algorithms present in the literature and we establish the convergence to the optimal solution.

Iterative hard thresholding (IHT) is a projected gradient descent algorithm, known to achieve state of the art performance for a wide range of structured estimation problems, such as sparse inference. In this work, we consider IHT as a solution to the problem of learning sparse discrete distributions. We study the hardness of using IHT on the space of measures. As a practical alternative, we propose a greedy approximate projection which simultaneously captures appropriate notions of sparsity in distributions, while satisfying the simplex constraint, and investigate the convergence behavior of the resulting procedure in various settings. Our results show, both in theory and practice, that IHT can achieve state of the art results for learning sparse distributions.

One of the most popular approaches to understanding feature effects of modern black box machine learning models are partial dependence plots (PDP). These plots are easy to understand but only able to visualize low order dependencies. The paper is about the question 'How much can we see?': A framework is developed to quantify the explainability of arbitrary machine learning models, i.e. up to what degree the visualization as given by a PDP is able to explain the predictions of the model. The result allows for a judgement whether an attempt to explain a black box model is sufficient or not.

Analysis of runner's data will often examine gait variables with reference to one or more gait events. Two such representative events are the initial contact and toe off events. These correspond respectively to the moments in time when the foot makes the initial contact with the ground and when the foot leaves the ground again. These variables are traditionally measured with a force plate or motion capture system in a lab setting. However, thanks to recent evolutions in wearable technology, the use of accelerometers has become commonplace for prolonged outdoor measurements. Previous research has developed heuristic methods to identify the initial contact and toe off timings based on minima, maxima and thresholds in the acceleration profiles. A significant flaw of these heuristic-based methods is that they are tailored to very specific acceleration profiles, providing no guidelines on how to handle deviant profiles. Therefore, we frame the problem as a structured prediction task and propose a machine learning approach for determining initial foot contact and toe off events from 3D tibial acceleration profiles. With mean absolute errors of 2 ms and 4 ms for respectively the initial contact and toe-off events, our method significantly outperforms the existing heuristic approaches.

Convolutional neural networks (CNNs) have been increasingly deployed to edge devices. Hence, many efforts have been made towards efficient CNN inference on resource-constrained platforms. This paper attempts to explore an orthogonal direction: how to conduct more energy-efficient training of CNNs, so as to enable on-device training? We strive to reduce the energy cost during training, by dropping unnecessary computations, from three complementary levels: stochastic mini-batch dropping on the data level; selective layer update on the model level; and sign prediction for low-cost, low-precision back-propagation, on the algorithm level. Extensive simulations and ablation studies, with real energy measurements from an FPGA board, confirm the superiority of our proposed strategies and demonstrate remarkable energy savings for training. For example, when training ResNet-74 on CIFAR-10, we achieve aggressive energy savings of >90% and >60%, while incurring a top-1 accuracy loss of only about 2% and 1.2%, respectively. When training ResNet-110 on CIFAR-100, an over 84% training energy saving is achieved without degrading inference accuracy.

Methods for automatic analysis of clinical data are usually targeted towards a specific modality and do not make use of all relevant data available. In the field of male human reproduction, clinical and biological data are not used to its fullest potential. Manual evaluation of a semen sample using a microscope is time-consuming and requires extensive training. Furthermore, the validity of manual semen analysis has been questioned due to limited reproducibility, and often high inter-personnel variation. The existing computer-aided sperm analyzer systems are not recommended for routine clinical use due to methodological challenges caused by the consistency of the semen sample. Thus, there is a need for an improved methodology. We use modern and classical machine learning techniques together with a dataset consisting of 85 videos of human semen samples and related participant data to automatically predict sperm motility. Used techniques include simple linear regression and more sophisticated methods using convolutional neural networks. Our results indicate that sperm motility prediction based on deep learning using sperm motility videos is rapid to perform and consistent. The algorithms performed worse when participant data was added. In conclusion, machine learning-based automatic analysis may become a valuable tool in male infertility investigation and research.

Universal probabilistic programming systems (PPSs) provide a powerful and expressive framework for specifying rich and complex probabilistic models. However, this expressiveness comes at the cost of substantially complicating the process of drawing inferences from the model. In particular, inference can become challenging when the support of the model varies between executions. Though general-purpose inference engines have been designed to operate in such settings, they are typically highly inefficient, often relying on proposing from the prior to make transitions. To address this, we introduce a new inference framework: Divide, Conquer, and Combine (DCC). DCC divides the program into separate straight-line sub-programs, each of which has a fixed support allowing more powerful inference algorithms to be run locally, before recombining their outputs in a principled fashion. We show how DCC can be implemented as an automated and general-purpose PPS inference engine, and empirically confirm that it can provide substantial performance improvements over previous approaches.