Neural Architecture Search (NAS) is an exciting new field which promises to be as much as a game-changer as Convolutional Neural Networks were in 2012. Despite many great works leading to substantial improvements on a variety of tasks, comparison between different methods is still very much an open issue. While most algorithms are tested on the same datasets, there is no shared experimental protocol followed by all. As such, and due to the under-use of ablation studies, there is a lack of clarity regarding why certain methods are more effective than others. Our first contribution is a benchmark of $8$ NAS methods on $5$ datasets. To overcome the hurdle of comparing methods with different search spaces, we propose using a method's relative improvement over the randomly sampled average architecture, which effectively removes advantages arising from expertly engineered search spaces or training protocols. Surprisingly, we find that many NAS techniques struggle to significantly beat the average architecture baseline. We perform further experiments with the commonly used DARTS search space in order to understand the contribution of each component in the NAS pipeline. These experiments highlight that: (i) the use of tricks in the evaluation protocol has a predominant impact on the reported performance of architectures; (ii) the cell-based search space has a very narrow accuracy range, such that the seed has a considerable impact on architecture rankings; (iii) the hand-designed macro-structure (cells) is more important than the searched micro-structure (operations); and (iv) the depth-gap is a real phenomenon, evidenced by the change in rankings between $8$ and $20$ cell architectures. To conclude, we suggest best practices, that we hope will prove useful for the community and help mitigate current NAS pitfalls. The code used is available at https://github.com/antoyang/NAS-Benchmark.

When presented with Out-of-Distribution (OOD) examples, deep neural networks yield confident, incorrect predictions. Detecting OOD examples is challenging, and the potential risks are high. In this paper, we propose to detect OOD examples by identifying inconsistencies between activity patterns and class predicted. We find that characterizing activity patterns by Gram matrices and identifying anomalies in gram matrix values can yield high OOD detection rates. We identify anomalies in the gram matrices by simply comparing each value with its respective range observed over the training data. Unlike many approaches, this can be used with any pre-trained softmax classifier and does not require access to OOD data for fine-tuning hyperparameters, nor does it require OOD access for inferring parameters. The method is applicable across a variety of architectures and vision datasets and, for the important and surprisingly hard task of detecting far-from-distribution out-of-distribution examples, it generally performs better than or equal to state-of-the-art OOD detection methods (including those that do assume access to OOD examples).

The recent increase in the availability of system condition monitoring data has lead to increases in the use of data-driven approaches for fault diagnostics. The accuracy of the fault detection and classification using these approaches is generally good when abundant labelled data on healthy and faulty system conditions exists and the diagnosis problem is formulated as a supervised learning task, i.e. supervised fault diagnosis. It is, however, relatively common in real situations that only a small fraction of the system condition monitoring data are labeled as healthy and the rest is unlabeled due to the uncertainty of the number and type of faults that may occur. In this case, supervised fault diagnosis performs poorly. Fault diagnosis with an unknown number and nature of faults is an open set learning problem where the knowledge of the faulty system is incomplete during training and the number and extent of the faults, of different types, can evolve during testing. In this paper, we propose to formulate the open set diagnostics problem as a semi-supervised learning problem and we demonstrate how it can be solved using a knowledge-induced learning approach with adaptive sampling variational autoencoders (KIL-AdaVAE) in combination with a one-class classifier. The fault detection and segmentation capability of the proposed method is demonstrated on a simulated case study using the Advanced Geared Turbofan 30000 (AGTF30) dynamical model under real flight conditions and induced faults of 17 fault types. The performance of the method is compared to the different learning strategies (supervised learning, supervised learning with embedding and semi-supervised learning) and deep learning algorithms. The results demonstrate that the proposed method is able to significantly outperform all other tested methods in terms of fault detection and fault segmentation.

Accurate short-term load forecasting is essential for efficient operation of the power sector. Predicting load at a fine granularity such as individual households or buildings is challenging due to higher volatility and uncertainty in the load. In aggregate loads such as at grids level, the inherent stochasticity and fluctuations are averaged-out, the problem becomes substantially easier. We propose an approach for short-term load forecasting at individual consumers (households) level, called {\em Forecasting using Matrix Factorization} (\textsc{FMF}). \textsc{FMF} does not use any consumers' demographic or activity patterns information. Therefore, it can be applied to any locality with the readily available smart meters and weather data. We perform extensive experiments on three benchmark datasets and demonstrate that \textsc{FMF} significantly outperforms the computationally expensive state-of-the-art methods for this problem. We achieve up to $26.5 \%$ and $24.4 \%$ improvement in \textsc{RMSE} over Regression Tree and Support Vector Machine, respectively and up to $36 \%$ and $73.2 \%$ improvement in \textsc{MAPE} over Random Forest and Long Short-Term Memory neural network, respectively.

Forecasting influenza-like illness (ILI) is of prime importance to epidemiologists and health-care providers. Early prediction of epidemic outbreaks plays a pivotal role in disease intervention and control. Most existing work has either limited long-term prediction performance or lacks a comprehensive ability to capture spatio-temporal dependencies in data. Accurate and early disease forecasting models would markedly improve both epidemic prevention and managing the onset of an epidemic. In this paper, we design a cross-location attention based graph neural network (Cola-GNN) for learning time series embeddings and location aware attentions. We propose a graph message passing framework to combine learned feature embeddings and an attention matrix to model disease propagation over time. We compare the proposed method with state-of-the-art statistical approaches and deep learning models on real-world epidemic-related datasets from United States and Japan. The proposed method shows strong predictive performance and leads to interpretable results for long-term epidemic predictions.

We introduce SLM Lab, a software framework for reproducible reinforcement learning (RL) research. SLM Lab implements a number of popular RL algorithms, provides synchronous and asynchronous parallel experiment execution, hyperparameter search, and result analysis. RL algorithms in SLM Lab are implemented in a modular way such that differences in algorithm performance can be confidently ascribed to differences between algorithms, not between implementations. In this work we present the design choices behind SLM Lab and use it to produce a comprehensive single-codebase RL algorithm benchmark. In addition, as a consequence of SLM Lab's modular design, we introduce and evaluate a discrete-action variant of the Soft Actor-Critic algorithm (Haarnoja et al., 2018) and a hybrid synchronous/asynchronous training method for RL agents.

Tesla's Neural Net continues to improve and become more advanced on a daily basis, but it appears that the electric car maker is making sure that it will evolve at an even faster rate in the future. A recent patent, for example, would allow Tesla's autonomous driving systems to work more efficiently, thanks to a new data pipeline focused on optimized image processing. Tesla's patent for "Data Pipeline and Deep Learning System for Autonomous Driving" was published on December 26. The idea behind the patent is to revolutionize and improve upon past deep learning systems that have been used for autonomous driving vehicles. In the past, these systems have used "captured sensor data" to retrieve information.

Despite the promising results obtained with [representations developed from Web image], the experiments demonstrate that object classification with real-life robotic data is far from being solved."

Artificial Neural Networks (ANNs) are the basis for everything that is happening in today's world. They might be smaller or bigger in size depending on the application but they are always there. Artificial Intelligence, Machine Learning, and Deep Learning are all powered by the ANNs. Although we have heard this term a lot, we have little to no knowledge of what they actually are. But you surely want to learn more about them in an easily understandable way; So today, let's talk Neural Networks!

On December 23, 2019, Yoshua Bengio and I debated the past and future of AI. Several thousand people tuned in, and tens of thousands watched afterward; ZDNet described it as a "historic event". Some people loved it, some hated it; I wished we had had more time. My slides and Yoshua's can be found here, as can a video (soon with improved sound) of the debate. Over the next days, I will try to articulate what I think was learned, and what remains open.