Numerous researchers recently applied empirical spectral analysis to the study of modern deep learning classifiers. We identify and discuss an important formal class/cross-class structure and show how it lies at the origin of the many visually striking features observed in deepnet spectra, some of which were reported in recent articles, others are unveiled here for the first time. These include spectral outliers, "spikes", and small but distinct continuous distributions, "bumps", often seen beyond the edge of a "main bulk". The significance of the cross-class structure is illustrated in three ways: (i) we prove the ratio of outliers to bulk in the spectrum of the Fisher information matrix is predictive of misclassification, in the context of multinomial logistic regression; (ii) we demonstrate how, gradually with depth, a network is able to separate class-distinctive information from class variability, all while orthogonalizing the class-distinctive information; and (iii) we propose a correction to KFAC, a well-known second-order optimization algorithm for training deepnets.
Recent work on Renyi Differential Privacy has shown the feasibility of applying differential privacy to deep learning tasks. Despite their promise, however, differentially private deep networks often lag far behind their non-private counterparts in accuracy, showing the need for more research in model architectures, optimizers, etc. One of the barriers to this expanded research is the training time -- often orders of magnitude larger than training non-private networks. The reason for this slowdown is a crucial privacy-related step called "per-example gradient clipping" whose naive implementation undoes the benefits of batch training with GPUs. By analyzing the back-propagation equations we derive new methods for per-example gradient clipping that are compatible with auto-differentiation (e.g., in PyTorch and TensorFlow) and provide better GPU utilization. Our implementation in PyTorch showed significant training speed-ups (by factors of 54x - 94x for training various models with batch sizes of 128). These techniques work for a variety of architectural choices including convolutional layers, recurrent networks, attention, residual blocks, etc.
The advent of the COVID-19 pandemic has instigated unprecedented changes in many countries around the globe, putting a significant burden on the health sectors, affecting the macro economic conditions, and altering social interactions amongst the population. In response, the academic community has produced multiple forecasting models, approaches and algorithms to best predict the different indicators of COVID-19, such as the number of confirmed infected cases. Yet, researchers had little to no historical information about the pandemic at their disposal in order to inform their forecasting methods. Our work studies the predictive performance of models at various stages of the pandemic to better understand their fundamental uncertainty and the impact of data availability on such forecasts. We use historical data of COVID-19 infections from 253 regions from the period of 22nd January 2020 until 22nd June 2020 to predict, through a rolling window backtesting framework, the cumulative number of infected cases for the next 7 and 28 days. We implement three simple models to track the root mean squared logarithmic error in this 6-month span, a baseline model that always predicts the last known value of the cumulative confirmed cases, a power growth model and an epidemiological model called SEIRD. Prediction errors are substantially higher in early stages of the pandemic, resulting from limited data. Throughout the course of the pandemic, errors regress slowly, but steadily. The more confirmed cases a country exhibits at any point in time, the lower the error in forecasting future confirmed cases. We emphasize the significance of having a rigorous backtesting framework to accurately assess the predictive power of such models at any point in time during the outbreak which in turn can be used to assign the right level of certainty to these forecasts and facilitate better planning.
Deep reinforcement learning (DRL) has reached super human levels in complex tasks like game solving (Go and autonomous driving). However, it remains an open question whether DRL can reach human level in applications to financial problems and in particular in detecting pattern crisis and consequently dis-investing. In this paper, we present an innovative DRL framework consisting in two sub-networks fed respectively with portfolio strategies past performances and standard deviations as well as additional contextual features. The second sub network plays an important role as it captures dependencies with common financial indicators features like risk aversion, economic surprise index and correlations between assets that allows taking into account context based information. We compare different network architectures either using layers of convolutions to reduce network's complexity or LSTM block to capture time dependency and whether previous allocations is important in the modeling. We also use adversarial training to make the final model more robust. Results on test set show this approach substantially over-performs traditional portfolio optimization methods like Markowitz and is able to detect and anticipate crisis like the current Covid one.
The Eulerian fluid simulation is an important HPC application. The neural network has been applied to accelerate it. The current methods that accelerate the fluid simulation with neural networks lack flexibility and generalization. In this paper, we tackle the above limitation and aim to enhance the applicability of neural networks in the Eulerian fluid simulation. We introduce Smartfluidnet, a framework that automates model generation and application. Given an existing neural network as input, Smartfluidnet generates multiple neural networks before the simulation to meet the execution time and simulation quality requirement. During the simulation, Smartfluidnet dynamically switches the neural networks to make the best efforts to reach the user requirement on simulation quality. Evaluating with 20,480 input problems, we show that Smartfluidnet achieves 1.46x and 590x speedup comparing with a state-of-the-art neural network model and the original fluid simulation respectively on an NVIDIA Titan X Pascal GPU, while providing better simulation quality than the state-of-the-art model.
The ability to know in advance the trend of running process instances, with respect to different features, such as the expected completion time, would allow business managers to timely counteract to undesired situations, in order to prevent losses. Therefore, the ability to accurately predict future features of running business process instances would be a very helpful aid when managing processes, especially under service level agreement constraints. However, making such accurate forecasts is not easy: many factors may influence the predicted features. Many approaches have been proposed to cope with this problem but all of them assume that the underling process is stationary. However, in real cases this assumption is not always true. In this work we present new methods for predicting the remaining time of running cases. In particular we propose a method, assuming process stationarity, which outperforms the state-of-the-art and two other methods which are able to make predictions even with non-stationary processes. We also describe an approach able to predict the full sequence of activities that a running case is going to take. All these methods are extensively evaluated on two real case studies.
Machine learning is nowadays the methodology of choice for flare forecasting and supervised techniques, in both their traditional and deep versions, are becoming the most frequently used ones for prediction in this area of space weather. Yet, machine learning has not been able so far to realize an operating warning system for flaring storms and the scientific literature of the last decade suggests that its performances in the prediction of intense solar flares are not optimal. The main difficulties related to forecasting solar flaring storms are probably two. First, most methods are conceived to provide probabilistic predictions and not to send binary yes/no indications on the consecutive occurrence of flares along an extended time range. Second, flaring storms are typically characterized by the explosion of high energy events, which are seldom recorded in the databases of space missions; as a consequence, supervised methods are trained on very imbalanced historical sets, which makes them particularly ineffective for the forecasting of intense flares. Yet, in this study we show that supervised machine learning could be utilized in a way to send timely warnings about the most violent and most unexpected flaring event of the last decade, and even to predict with some accuracy the energy budget daily released by magnetic reconnection during the whole time course of the storm. Further, we show that the combination of sparsity-enhancing machine learning and feature ranking could allow the identification of the prominent role that energy played as an Active Region property in the forecasting process.
This paper proposes an out-of-sample extension framework for a global manifold learning algorithm (Isomap) that uses temporal information in out-of-sample points in order to make the embedding more robust to noise and artifacts. Given a set of noise-free training data and its embedding, the proposed framework extends the embedding for a noisy time series. This is achieved by adding a spatio-temporal compactness term to the optimization objective of the embedding. To the best of our knowledge, this is the first method for out-of-sample extension of manifold embeddings that leverages timing information available for the extension set. Experimental results demonstrate that our out-of-sample extension algorithm renders a more robust and accurate embedding of sequentially ordered image data in the presence of various noise and artifacts when compared to other timing-aware embeddings. Additionally, we show that an out-of-sample extension framework based on the proposed algorithm outperforms the state of the art in eye-gaze estimation.
In this work, we are interested in the large graph similarity computation problem, which is one of the most important graph-based problems. Traditional techniques to compute the exact or approximate values of Graph Edit Distance (GED) and Maximum Common Subgraph (MCS) require at least polynomial time complexity over node numbers thus are not able to handle this problem when the numbers of nodes are large. Recently the develop of deep learning techniques provide a promising solution for this problem by training a network which is able to encode graphs to feature vectors and then compute similarity based on feature vectors. However, when we look into these techniques and classify them to embedding models and matching models, problems arise. Embedding models can be quite fast but perform poorly due to the lack of interaction across graphs while matching models involve this for much better performance but satisfy far more on time consumption. Similar to the process of large biological molecular identification, where we first maps the whole molecular to molecular groups and then identify them based on the "abstracted smaller molecular", the feature aggregation across two whole graphs is always redundant, especially when the number of nodes is large. Thus we here present a novel framework for large graph similarity computation problem. We first embed and coarsen the large graphs to "abstracted smaller graphs" with denser local topology, similar to molecular groups in biological concept. Then we aggregate both the internal features in "abstracted smaller graphs" and external features across "abstracted smaller graph pair", leading to feature vectors for each graph, with which we calculate the final similarity score. Experiments demonstrate that our proposed framework outperforms state-of-the-art methods in graph similarity computation tasks and has significant improvement in time efficiency.
In this paper we propose a convolutional autoencoder to address the problem of motion infilling for 3D human motion data. Given a start and end sequence, motion infilling aims to complete the missing gap in between, such that the filled in poses plausibly forecast the start sequence and naturally transition into the end sequence. To this end, we propose a single, end-to-end trainable convolutional autoencoder. We show that a single model can be used to create natural transitions between different types of activities. Furthermore, our method is not only able to fill in entire missing frames, but it can also be used to complete gaps where partial poses are available (e.g. from end effectors), or to clean up other forms of noise (e.g. Gaussian). Also, the model can fill in an arbitrary number of gaps that potentially vary in length. In addition, no further post-processing on the model's outputs is necessary such as smoothing or closing discontinuities at the end of the gap. At the heart of our approach lies the idea to cast motion infilling as an inpainting problem and to train a convolutional de-noising autoencoder on image-like representations of motion sequences. At training time, blocks of columns are removed from such images and we ask the model to fill in the gaps. We demonstrate the versatility of the approach via a number of complex motion sequences and report on thorough evaluations performed to better understand the capabilities and limitations of the proposed approach.