Entity linking (EL), the task of disambiguating mentions in text by linking them to entities in a knowledge graph, is crucial for text understanding, question answering or conversational systems. Entity linking on short text (e.g., single sentence or question) poses particular challenges due to limited context. While prior approaches use either heuristics or black-box neural methods, here we propose LNN-EL, a neuro-symbolic approach that combines the advantages of using interpretable rules based on first-order logic with the performance of neural learning. Even though constrained to using rules, LNN-EL performs competitively against SotA black-box neural approaches, with the added benefits of extensibility and transferability. In particular, we show that we can easily blend existing rule templates given by a human expert, with multiple types of features (priors, BERT encodings, box embeddings, etc), and even scores resulting from previous EL methods, thus improving on such methods. For instance, on the LC-QuAD-1.0 dataset, we show more than $4$\% increase in F1 score over previous SotA. Finally, we show that the inductive bias offered by using logic results in learned rules that transfer well across datasets, even without fine tuning, while maintaining high accuracy.

Detecting what emotions are expressed in text is a well-studied problem in natural language processing. However, research on finer grained emotion analysis such as what causes an emotion is still in its infancy. We present solutions that tackle both emotion recognition and emotion cause detection in a joint fashion. Considering that common-sense knowledge plays an important role in understanding implicitly expressed emotions and the reasons for those emotions, we propose novel methods that combine common-sense knowledge via adapted knowledge models with multi-task learning to perform joint emotion classification and emotion cause tagging. We show performance improvement on both tasks when including common-sense reasoning and a multitask framework. We provide a thorough analysis to gain insights into model performance.

This paper investigates two techniques for developing efficient self-supervised vision transformers (EsViT) for visual representation learning. First, we show through a comprehensive empirical study that multi-stage architectures with sparse self-attentions can significantly reduce modeling complexity but with a cost of losing the ability to capture fine-grained correspondences between image regions. Second, we propose a new pre-training task of region matching which allows the model to capture fine-grained region dependencies and as a result significantly improves the quality of the learned vision representations. Our results show that combining the two techniques, EsViT achieves 81.3% top-1 on the ImageNet linear probe evaluation, outperforming prior arts with around an order magnitude of higher throughput. When transferring to downstream linear classification tasks, EsViT outperforms its supervised counterpart on 17 out of 18 datasets. The code and models will be publicly available.

The performance of a reinforcement learning (RL) system depends on the computational architecture used to approximate a value function. Deep learning methods provide both optimization techniques and architectures for approximating nonlinear functions from noisy, high-dimensional observations. However, prevailing optimization techniques are not designed for strictly-incremental online updates. Nor are standard architectures designed for observations with an a priori unknown structure: for example, light sensors randomly dispersed in space. This paper proposes an online RL prediction algorithm with an adaptive architecture that efficiently finds useful nonlinear features. The algorithm is evaluated in a spatial domain with high-dimensional, stochastic observations. The algorithm outperforms non-adaptive baseline architectures and approaches the performance of an architecture given side-channel information. These results are a step towards scalable RL algorithms for more general problems, where the observation structure is not available.

Neural networks are increasingly used in environmental science applications. Furthermore, neural network models are trained by minimizing a loss function, and it is crucial to choose the loss function very carefully for environmental science applications, as it determines what exactly is being optimized. Standard loss functions do not cover all the needs of the environmental sciences, which makes it important for scientists to be able to develop their own custom loss functions so that they can implement many of the classic performance measures already developed in environmental science, including measures developed for spatial model verification. However, there are very few resources available that cover the basics of custom loss function development comprehensively, and to the best of our knowledge none that focus on the needs of environmental scientists. This document seeks to fill this gap by providing a guide on how to write custom loss functions targeted toward environmental science applications. Topics include the basics of writing custom loss functions, common pitfalls, functions to use in loss functions, examples such as fractions skill score as loss function, how to incorporate physical constraints, discrete and soft discretization, and concepts such as focal, robust, and adaptive loss. While examples are currently provided in this guide for Python with Keras and the TensorFlow backend, the basic concepts also apply to other environments, such as Python with PyTorch. Similarly, while the sample loss functions provided here are from meteorology, these are just examples of how to create custom loss functions. Other fields in the environmental sciences have very similar needs for custom loss functions, e.g., for evaluating spatial forecasts effectively, and the concepts discussed here can be applied there as well. All code samples are provided in a GitHub repository.

Machine learning technology has the potential to dramatically optimise event generation and simulations. We continue to investigate the use of neural networks to approximate matrix elements for high-multiplicity scattering processes. We focus on the case of loop-induced diphoton production through gluon fusion and develop a realistic simulation method that can be applied to hadron collider observables. Neural networks are trained using the one-loop amplitudes implemented in the NJet C++ library and interfaced to the Sherpa Monte Carlo event generator where we perform a detailed study for $2\to3$ and $2\to4$ scattering problems. We also consider how the trained networks perform when varying the kinematic cuts effecting the phase space and the reliability of the neural network simulations.

Class-imbalanced data, in which some classes contain far more samples than others, is ubiquitous in real-world applications. Standard techniques for handling class-imbalance usually work by training on a re-weighted loss or on re-balanced data. Unfortunately, training overparameterized neural networks on such objectives causes rapid memorization of minority class data. To avoid this trap, we harness meta-learning, which uses both an ''outer-loop'' and an ''inner-loop'' loss, each of which may be balanced using different strategies. We evaluate our method, MetaBalance, on image classification, credit-card fraud detection, loan default prediction, and facial recognition tasks with severely imbalanced data, and we find that MetaBalance outperforms a wide array of popular re-sampling strategies.

Deep Convolutional Neural Networks are opaque machine learning solutions. Opaque in the sense that the model state itself is neither self evident nor human interpretable. This has lead to a primarily trial and error driven approach for development, that relies on the comparison of abstract, model agnostic performance metrics like accuracy as a measure for classification performance, the number of parameters as a measure of capacity and the FLOPs per forward pass as a metric for computational efficiency Tan et al. (2018); Tan and Le (2019); He et al. (2015). To move towards a more efficient, less trial and error based design process, a deeper understanding of the model's state is required. This understanding does not have to be necessarily complete with regards to fully understanding the relation of the input and output of the model. The comparative analysis methods based on SVCCA Raghu et al. (2017) are good examples of such a non-holistic approaches. The information extracted from the model by using SVCCA is highly aggregated but allows for useful insights into the converged model. Logistic Regression Probes Alain and Bengio (2018) and saturation Richter et al. (2020) aggregate a single layer to a number, which allows for easy and intuitive analysis, similar to measuring with a thermometer. While logistic regression probes measure the intermediate solution quality very directly by training logistic regressions on the output of a layer, saturation is more task agnostic.

Path representations are critical in a variety of transportation applications, such as estimating path ranking in path recommendation systems and estimating path travel time in navigation systems. Existing studies often learn task-specific path representations in a supervised manner, which require a large amount of labeled training data and generalize poorly to other tasks. We propose an unsupervised learning framework Path InfoMax (PIM) to learn generic path representations that work for different downstream tasks. We first propose a curriculum negative sampling method, for each input path, to generate a small amount of negative paths, by following the principles of curriculum learning. Next, \emph{PIM} employs mutual information maximization to learn path representations from both a global and a local view. In the global view, PIM distinguishes the representations of the input paths from those of the negative paths. In the local view, \emph{PIM} distinguishes the input path representations from the representations of the nodes that appear only in the negative paths. This enables the learned path representations to encode both global and local information at different scales. Extensive experiments on two downstream tasks, ranking score estimation and travel time estimation, using two road network datasets suggest that PIM significantly outperforms other unsupervised methods and is also able to be used as a pre-training method to enhance supervised path representation learning.

Recently, there has been an increasing interest in neural speech synthesis. While the deep neural network achieves the state-of-the-art result in text-to-speech (TTS) tasks, how to generate a more emotional and more expressive speech is becoming a new challenge to researchers due to the scarcity of high-quality emotion speech dataset and the lack of advanced emotional TTS model. In this paper, we first briefly introduce and publicly release a Mandarin emotion speech dataset including 9,724 samples with audio files and its emotion human-labeled annotation. After that, we propose a simple but efficient architecture for emotional speech synthesis called EMSpeech. Unlike those models which need additional reference audio as input, our model could predict emotion labels just from the input text and generate more expressive speech conditioned on the emotion embedding. In the experiment phase, we first validate the effectiveness of our dataset by an emotion classification task. Then we train our model on the proposed dataset and conduct a series of subjective evaluations. Finally, by showing a comparable performance in the emotional speech synthesis task, we successfully demonstrate the ability of the proposed model.