As machine learning (ML) models find wide-spread application, there is a growing demand for interpretability: access to tools that help people see why the model made its decision. There are still many obstacles towards achieving this goal though, particularly in the context of deep learning. These obstacles stem from the scale of modern deep networks, as well as the complexity of even defining and assessing the (often context-dependent) desiderata of interpretability. Existing work on deep network interpretability has largely approached this problem from two perspectives. The first one seeks to uncover the concepts associated with specific neurons in the network, for example through visualization [Yos 15] or semantic labeling [Bau 17].

Rationalizing which parts of a molecule drive the predictions of a molecular graph convolutional neural network (GCNN) can be difficult. To help, we propose two simple regularization techniques to apply during the training of GCNNs: Batch Representation Orthonormalization (BRO) and Gini regularization. BRO, inspired by molecular orbital theory, encourages graph convolution operations to generate orthonormal node embeddings. Gini regularization is applied to the weights of the output layer and constrains the number of dimensions the model can use to make predictions. We show that Gini and BRO regularization can improve the accuracy of state-of-the-art GCNN attribution methods on artificial benchmark datasets. In a real-world setting, we demonstrate that medicinal chemists significantly prefer explanations extracted from regularized models. While we only study these regularizers in the context of GCNNs, both can be applied to other types of neural networks

We propose two novel loss functions based on Jensen-Shannon divergence for learning under label noise. Following the work of Ghosh et al. (2017), we argue about their theoretical robustness. Furthermore, we reveal several other desirable properties by drawing informative connections to various loss functions, e.g., cross entropy, mean absolute error, generalized cross entropy, symmetric cross entropy, label smoothing, and most importantly consistency regularization. We conduct extensive and systematic experiments using both synthetic (CIFAR) and real (WebVision) noise and demonstrate significant and consistent improvements over other loss functions. Also, we conduct several informative side experiments that highlight the different theoretical properties.

Marginal-likelihood based model-selection, even though promising, is rarely used in deep learning due to estimation difficulties. Instead, most approaches rely on validation data, which may not be readily available. In this work, we present a scalable marginal-likelihood estimation method to select both the hyperparameters and network architecture based on the training data alone. Some hyperparameters can be estimated online during training, simplifying the procedure. Our marginal-likelihood estimate is based on Laplace's method and Gauss-Newton approximations to the Hessian, and it outperforms cross-validation and manual-tuning on standard regression and image classification datasets, especially in terms of calibration and out-of-distribution detection. Our work shows that marginal likelihoods can improve generalization and be useful when validation data is unavailable (e.g., in nonstationary settings).

This perspective piece came about through the Generative Adversarial Collaboration (GAC) series of workshops organized by the Computational Cognitive Neuroscience (CCN) conference in 2020. We brought together a number of experts from the field of theoretical neuroscience to debate emerging issues in our understanding of how learning is implemented in biological recurrent neural networks. Here, we will give a brief review of the common assumptions about biological learning and the corresponding findings from experimental neuroscience and contrast them with the efficiency of gradient-based learning in recurrent neural networks commonly used in artificial intelligence. We will then outline the key issues discussed in the workshop: synaptic plasticity, neural circuits, theory-experiment divide, and objective functions. Finally, we conclude with recommendations for both theoretical and experimental neuroscientists when designing new studies that could help to bring clarity to these issues.

Neuro-Symbolic Artificial Intelligence -- the combination of symbolic methods with methods that are based on artificial neural networks -- has a long-standing history. In this article, we provide a structured overview of current trends, by means of categorizing recent publications from key conferences. The article is meant to serve as a convenient starting point for research on the general topic.

The field received significant interest in the previous decade, mainly due to advances in automated machine learning (ML) and deep learning (DL). They learn useful patterns from a large amount of data and keep the acquired knowledge as model structures and parameters that can be further applied to make predictions by interpreting unseen data [1]. These models are either a set of elements or features that contribute when making decisions (in ML) or are organized into several layers of abstraction (such as neural networks) for general and specific interpretation tasks (in DL). Healthcare provision and medicine are one of the most significant challenges of AI due to being the pillars for a global society and the necessity of providing higher-quality assistance to the healthcare workforce [2]. An emerging and expanding domain for for application of AI is mental health. Readily available and ubiquitous devices and applications enable the provision of flexible mental care - on-demand, at any time, both at healthcare facilities and at home.

This paper describes our contribution to SemEval-2020 Task 7: Assessing Humor in Edited News Headlines. Here we present a method based on a deep neural network. In recent years, quite some attention has been devoted to humor production and perception. Our team KdeHumor employs recurrent neural network models including Bi-Directional LSTMs (BiLSTMs). Moreover, we utilize the state-of-the-art pre-trained sentence embedding techniques. We analyze the performance of our method and demonstrate the contribution of each component of our architecture.

Humour detection from sentences has been an interesting and challenging task in the last few years. In attempts to highlight humour detection, most research was conducted using traditional approaches of embedding, e.g., Word2Vec or Glove. Recently BERT sentence embedding has also been used for this task. In this paper, we propose a framework for humour detection in short texts taken from news headlines. Our proposed framework (IBEN) attempts to extract information from written text via the use of different layers of BERT. After several trials, weights were assigned to different layers of the BERT model. The extracted information was then sent to a Bi-GRU neural network as an embedding matrix. We utilized the properties of some external embedding models. A multi-kernel convolution in our neural network was also employed to extract higher-level sentence representations. This framework performed very well on the task of humour detection.

Artificial Intelligence (AI) and Deep Learning (DL) algorithms are currently applied to a wide range of products and solutions. DL training jobs are highly resource demanding and they experience great benefits when exploiting AI accelerators (e.g., GPUs). However, the effective management of GPU-powered clusters comes with great challenges. Among these, efficient scheduling and resource allocation solutions are crucial to maximize performance and minimize Data Centers operational costs. In this paper we propose ANDREAS, an advanced scheduling solution that tackles these problems jointly, aiming at optimizing DL training runtime workloads and their energy consumption in accelerated clusters. Experiments based on simulation demostrate that we can achieve a cost reduction between 30 and 62% on average with respect to first-principle methods while the validation on a real cluster shows a worst case deviation below 13% between actual and predicted costs, proving the effectiveness of ANDREAS solution in practical scenarios.