Transfer Learning in AI is a method where a model is developed for a specific task, which is used as the initial steps for another model for other tasks. Deep Convolutional Neural Networks in deep learning take an hour or day to train the mode if the dataset we are playing is vast. The approach is we reuse the weights of the pre-trained model, which was trained for some standard Computer Vision datasets such as Image classification (Image Net). Extensive deep Convolutional networks for large-scale image classification are available in Keras, which we can directly import and can be used with their pre-trained weights. Let's now understand how to use VGG16 pre-trained on 10,000 categories(Image Net) for the Distracted driver Detection dataset.

Creating complex data and analysis pipelines has never been easier. You'll be inundated with tutorials online. You can learn the language at every turn. Keeping track of it all is not so easy. Learning the programming basics is easy, but keeping track of the technological possibilities only grows with experience. We present you Awesome Python Data Science libraries and frameworks for free that you should know.

McGill study aims to make social media contributions more useful to crisis managers Can combining deep learning (DL)— a subfield of artificial intelligence— with social network analysis (SNA), make social media contributions about extreme weather events a useful tool for crisis managers, first responders and government scientists? An interdisciplinary team of McGill researchers has brought these tools to the forefront in an effort to understand and manage extreme weather events. The researchers found that by using a noise reduction mechanism, valuable information could be filtered from social media to better assess trouble spots and assess users’ reactions vis-à-vis extreme weather events. The results of the study are published in the Journal of Contingencies and Crisis Management. Diving into a sea of information “We reduced the noise by finding out who was being listened to, and which were authoritative sources,” explains Renee Sieber, Associate Professor in McGill’s Department of Geography and lead author of this study. “This ability is important because it is quite difficult to assess the validity of the information shared by Twitter users.” The team based their study on Twitter data from the March 2019 Nebraska floods in the United States, which caused over $1 billion in damage and widespread evacuations of residents. In total, over 1,200 tweets were analyzed and classified. “Social network analysis can identify where ​people get their information during an extreme weather event. Deep learning allows us to better understand the content ​ of this information by classifying thousands of tweets into fixed categories, for example, ‘infrastructure and utilities damage’ or ‘sympathy and emotional support’,” says Sieber. The researchers then introduced a two-tiered DL classification model – a first in terms of integrating these methods in a way that could be useful to crisis managers. The study highlighted some issues regarding the use of social media analysis for this purpose, notably its failure to note that events are far more contextual than expected by labelled datasets, such as the CrisisNLP, and the lack of a universal language to categorize terms related to crisis management. The preliminary exploration performed by the researchers also found that a celebrity call out was featured prominently – this was indeed the case for the 2019 Nebraska floods, where a tweet from pop singer Justin Timberlake was shared by a large number of users, though it did not prove to be of use for crisis managers. “Our findings tell us that information content varies between different types of events, contrary to the belief that there is a universal language to categorize crisis management; this limits the use of labelled datasets on just a few types of events, as search terms may change from one event to another.” “The vast amount of social media data the public contributes about weather suggests it can provide critical information in crises, such as snowstorms, floods, and ice storms. We are currently exploring transferring this model to different types of weather crises and addressing the shortcomings of existing supervised approaches by combining these with other methods,” says Sieber. About this study “Using deep learning and social network analysis to understand and manage extreme flooding” by Renee Sieber and al. was published in the Journal of Contingencies and Crisis Management. This study was funded by Environment Canada. About McGill University Founded in Montreal, Quebec, in 1821, McGill University is Canada’s top ranked medical doctoral university. McGill is consistently ranked as one of the top universities, both nationally and internationally. It is a world-renowned institution of higher learning with research activities spanning two campuses, 11 faculties, 13 professional schools, 300 programs of study and over 40,000 students, including more than 10,200 graduate students. McGill attracts students from over 150 countries around the world, its 12,800 international students making up 31% per cent of the student body. Over half of McGill students claim a first language other than English, including approximately 19% of our students who say French is their mother tongue.

Researchers from Kent have developed a computer algorithm that can identify differences in cancer cell lines based on microscopic images, a unique development towards ending misidentification of cells in laboratories. Cancer cell lines are cells isolated and grown as cell cultures in laboratories for study and developing anti-cancer drugs. However, many cell lines are misidentified after being swapped or contaminated with others, meaning many researchers may work with incorrect cells. This has been a persistent problem since work with cancer cell lines began. Short tandem repeat (STR) analysis is commonly used to identify cancer cell lines, but is expensive and time-consuming.

The unstructured sparsity after pruning poses a challenge to the efficient implementation of deep learning models in existing regular architectures like systolic arrays. On the other hand, coarse-grained structured pruning is suitable for implementation in regular architectures but tends to have higher accuracy loss than unstructured pruning when the pruned models are of the same size. In this work, we propose a model compression method based on a novel weight permutation scheme to fully exploit the fine-grained weight sparsity in the hardware design. Through permutation, the optimal arrangement of the weight matrix is obtained, and the sparse weight matrix is further compressed to a small and dense format to make full use of the hardware resources. Two pruning granularities are explored. In addition to the unstructured weight pruning, we also propose a more fine-grained subword-level pruning to further improve the compression performance. Compared to the state-of-the-art works, the matrix compression rate is significantly improved from 5.88x to 14.13x. As a result, the throughput and energy efficiency are improved by 2.75 and 1.86 times, respectively.

It is well known that numerical simulations of high-speed reacting flows, in the framework of state-to-state formulations, are the most detailed but also often prohibitively computationally expensive. In this work, we start to investigate the possibilities offered by the use of machine learning methods for state-to-state approaches to alleviate such burden. In this regard, several tasks have been identified. Firstly, we assessed the potential of state-of-the-art data-driven regression models based on machine learning to predict the relaxation source terms which appear in the right-hand side of the state-to-state Euler system of equations for a one-dimensional reacting flow of a N$_2$/N binary mixture behind a plane shock wave. It is found that, by appropriately choosing the regressor and opportunely tuning its hyperparameters, it is possible to achieve accurate predictions compared to the full-scale state-to-state simulation in significantly shorter times. Secondly, we investigated different strategies to speed-up our in-house state-to-state solver by coupling it with the best-performing pre-trained machine learning algorithm. The embedding of machine learning methods into ordinary differential equations solvers may offer a speed-up of several orders of magnitude but some care should be paid for how and where such coupling is realized. Performances are found to be strongly dependent on the mutual nature of the interfaced codes. Finally, we aimed at inferring the full solution of the state-to-state Euler system of equations by means of a deep neural network completely by-passing the use of the state-to-state solver while relying only on data. Promising results suggest that deep neural networks appear to be a viable technology also for these tasks.

With the increase of available time series data, predicting their class labels has been one of the most important challenges in a wide range of disciplines. Recent studies on time series classification show that convolutional neural networks (CNN) achieved the state-of-the-art performance as a single classifier. In this work, pointing out that the global pooling layer that is usually adopted by existing CNN classifiers discards the temporal information of high-level features, we present a dynamic temporal pooling (DTP) technique that reduces the temporal size of hidden representations by aggregating the features at the segment-level. For the partition of a whole series into multiple segments, we utilize dynamic time warping (DTW) to align each time point in a temporal order with the prototypical features of the segments, which can be optimized simultaneously with the network parameters of CNN classifiers. The DTP layer combined with a fully-connected layer helps to extract further discriminative features considering their temporal position within an input time series. Extensive experiments on both univariate and multivariate time series datasets show that our proposed pooling significantly improves the classification performance.

Auto-regressive sequence-to-sequence models with attention mechanisms have achieved state-of-the-art performance in various tasks including Text-To-Speech (TTS) and Neural Machine Translation (NMT). The standard training approach, teacher forcing, guides a model with the reference output history. At inference stage, the generated output history must be used. This mismatch can impact performance. However, it is highly challenging to train the model using the generated output. Several approaches have been proposed to address this problem, normally by selectively using the generated output history. To make training stable, these approaches often require a heuristic schedule or an auxiliary classifier. This paper introduces attention forcing for NMT. This approach guides the model with the generated output history and reference attention, and can reduce the training-inference mismatch without a schedule or a classifier. Attention forcing has been successful in TTS, but its application to NMT is more challenging, due to the discrete and multi-modal nature of the output space. To tackle this problem, this paper adds a selection scheme to vanilla attention forcing, which automatically selects a suitable training approach for each pair of training data. Experiments show that attention forcing can improve the overall translation quality and the diversity of the translations.

Deep neural networks achieve high prediction accuracy when the train and test distributions coincide. However, in practice various types of corruptions can deviate from this setup and performance can be heavily degraded. There have been only a few methods to address generalization in presence of unexpected domain shifts observed during deployment. In this paper, a misclassification-aware Gaussian smoothing approach is presented to improve the robustness of image classifiers against a variety of corruptions while maintaining clean accuracy. The intuition behind our proposed misclassification-aware objective is revealed through bounds on the local loss deviation in the small-noise regime. When our method is coupled with additional data augmentations, it is empirically shown to improve upon the state-of-the-art in robustness and uncertainty calibration on several image classification tasks.

This paper addresses distributional offline continuous-time reinforcement learning (DOCTR-L) with stochastic policies for high-dimensional optimal control. A soft distributional version of the classical Hamilton-Jacobi-Bellman (HJB) equation is given by a semilinear partial differential equation (PDE). This `soft HJB equation' can be learned from offline data without assuming that the latter correspond to a previous optimal or near-optimal policy. A data-driven solution of the soft HJB equation uses methods of Neural PDEs and Physics-Informed Neural Networks developed in the field of Scientific Machine Learning (SciML). The suggested approach, dubbed `SciPhy RL', thus reduces DOCTR-L to solving neural PDEs from data. Our algorithm called Deep DOCTR-L converts offline high-dimensional data into an optimal policy in one step by reducing it to supervised learning, instead of relying on value iteration or policy iteration methods. The method enables a computable approach to the quality control of obtained policies in terms of both their expected returns and uncertainties about their values.