Saliency methods help to make deep neural network predictions more interpretable by identifying particular features, such as pixels in an image, that contribute most strongly to the network's prediction. Unfortunately, recent evidence suggests that many saliency methods perform poorly when gradients are saturated or in the presence of strong inter-feature dependence or noise injected by an adversarial attack. In this work, we propose to infer robust saliency scores by integrating the saliency scores of a set of decoys with a novel decoy-enhanced saliency score, in which the decoys are generated by either solving an optimization problem or blurring the original input. We theoretically analyze that our method compensates for gradient saturation and considers joint activation patterns of pixels. We also apply our method to three different CNNs---VGGNet, AlexNet, and ResNet trained on ImageNet data set. The empirical results show both qualitatively and quantitatively that our method outperforms raw scores produced by three existing saliency methods, even in the presence of adversarial attacks.

Data-driven approaches, most prominently deep learning, have become powerful tools for prediction in many domains. A natural question to ask is whether data-driven methods could also be used for numerical weather prediction. First studies show promise but the lack of a common dataset and evaluation metrics make inter-comparison between studies difficult. Here we present a benchmark dataset for data-driven medium-range weather forecasting, a topic of high scientific interest for atmospheric and computer scientists alike. We provide data derived from the ERA5 archive that has been processed to facilitate the use in machine learning models. We propose a simple and clear evaluation metric which will enable a direct comparison between different methods. Further, we provide baseline scores from simple linear regression techniques, deep learning models as well as purely physical forecasting models. All data is publicly available and the companion code is reproducible with tutorials for getting started. We hope that this dataset will accelerate research in data-driven weather forecasting.

Designing a deep neural network (DNN) with good generalization capability is a complex process especially when the weights are severely quantized. Model averaging is a promising approach for achieving the good generalization capability of DNNs, especially when the loss surface for training contains many sharp minima. We present a new quantized neural network optimization approach, stochastic quantized weight averaging (SQWA), to design low-precision DNNs with good generalization capability using model averaging. The proposed approach includes (1) floating-point model training, (2) direct quantization of weights, (3) capturing multiple low-precision models during retraining with cyclical learning rates, (4) averaging the captured models, and (5) re-quantizing the averaged model and fine-tuning it with low-learning rates. Additionally, we present a loss-visualization technique on the quantized weight domain to clearly elucidate the behavior of the proposed method. Visualization results indicate that a quantized DNN (QDNN) optimized with the proposed approach is located near the center of the flat minimum in the loss surface. With SQWA training, we achieved state-of-the-art results for 2-bit QDNNs on CIFAR-100 and ImageNet datasets. Although we only employed a uniform quantization scheme for the sake of implementation in VLSI or low-precision neural processing units, the performance achieved exceeded those of previous studies employing non-uniform quantization.

It's easy to understand the importance of NLP given the number of applications for it--question-and-answer (Q&A) systems, translation of text from one language to another, automatic summarization (of long texts into short summaries), grammar analysis and recommendation, sentiment analysis, and much more. This technology is even more important today given the massive amount of unstructured data generated daily in the context of news, social media, scientific and technical papers, and the variety of other sources in our connected world. Today, when we ask Alexa or Siri a question, we don't think about the complexity involved in recognizing speech, understanding the meaning of the question, and ultimately providing a response. Recent advances in state-of-the-art NLP models, BERT, and BERT's lighter successor ALBERT from Google is setting new benchmarks in the industry and allowing researchers to increase training speed of the models. In the mid-1950s, IBM sparked tremendous excitement for language understanding through what was called the Georgetown experiment, a joint development project between IBM and Georgetown University.

Researchers from the RIKEN Center for Advanced Intelligence Project (AIP) in Japan have shown that a deep-learning algorithm can be used to extract interpretable features from annotation-free histopathology images from prostate cancer patients. Their framework outperformed the prediction of biochemical recurrence using conventional, Gleason Score-based methods (Nature Commun. Prostate cancer is the second most common cancer affecting men worldwide, with an incidence rate of 13.5%, according to the World Health Organization. The extracted samples of tissue are examined under a microscope and, if cancerous cells are found, divided into risk groups assigned through the Gleason Score. This grading system is considered the gold standard in cancer medicine, as it determines the aggressiveness of prostate cancer and helps doctors establish the right course of treatment.

Facebook's AI team has been extra thirsty for robotics lately. Over the course of the past year, Facebook has expanded its robotics operations around the world and taught hexapod robots to walk. Last week, Facebook AI released the Replica photorealistic training data set and Habitat, a simulation engine for embodied AI, like robots. As part of that same effort, Facebook AI today introduced open source robotics framework PyRobot. Created in collaboration with Carnegie Mellon University researchers, PyRobot can run deep learning models trained by Facebook's machine learning framework PyTorch.

The 20th century turned out to be an era of exponential growth in the field of machine learning. The 3000-year-old ancient game of'Go' that computer scientists predicted will take another decade to crack was made possible by Google Brain teams AlphaGo AI, defeating multiple-time world champion Lee Sudol. And, by the way, this Chinese game has more combinations than predicted atoms in the universe or, in short, this game can't be won just by running through all the possible moves, what IBM Blue did in 1997, defeating world champion Gary Kasparov. Research communities are thriving in ML, from 100 papers submitted annually 10 years ago, to 100 per day in 2019 on arXiv alone. But, keeping everything aside, the point is that ML is highly math-intensive. While libraries like TensorFlow and PyTorch have made a significant contribution in making ML reachable to all the developers out there, we still have a steep learning curve to know how to create models, train them, and save it to later use it for our tasks.

A few days ago, Amazon announced the release of AutoGloun, a new toolkit that simplifies the creation of deep learning models with just a few lines of code. The release marks Amazon's entrance in the ultra-competitive Automated machine learning(AutoML) space which is becoming one of the hottest trends for enterprise machine learning platforms. With some many news around the AutoML ecosystem, sometimes it becomes hard to differentiate signal from noise. Today, I would like to explore some of the most innovative AutoML stacks in the market that don't receive that much publicity. AutoML is becoming one of the most popular topics in modern data science applications.

"Just like software, and the Internet from previous decades, public cloud and now AI are the megatrends of our generation." Artificial intelligence and machine learning (AI/ML) is driving breakthrough developments across industries such as Healthcare, Energy, Logistics, and more. Heliogen is using AI to optimize the next generation of solar technology to power energy intensive processes such as manufacturing steel which in the past was only possible with fossil fuels. Another example is Boston Dynamics' HANDLE – an agile mobile robot that uses deep learning to autonomously unload trucks and move boxes in warehouses. If someone tells you that AI/ML is hype, remind them that cloud computing was once called hype.

"Just like software, and the Internet from previous decades, public cloud and now AI are the megatrends of our generation." Artificial intelligence and machine learning (AI/ML) is driving breakthrough developments across industries such as Healthcare, Energy, Logistics, and more. Heliogen is using AI to optimize the next generation of solar technology to power energy intensive processes such as manufacturing steel which in the past was only possible with fossil fuels. Another example is Boston Dynamics' HANDLE – an agile mobile robot that uses deep learning to autonomously unload trucks and move boxes in warehouses. If someone tells you that AI/ML is hype, remind them that cloud computing was once called hype.