Recent advances in deep learning have facilitated the demand of neural models for real applications. In practice, these applications often need to be deployed with limited resources while keeping high accuracy. This paper touches the core of neural models in NLP, word embeddings, and presents a new embedding distillation framework that remarkably reduces the dimension of word embeddings without compromising accuracy. A novel distillation ensemble approach is also proposed that trains a high-efficient student model using multiple teacher models. In our approach, the teacher models play roles only during training such that the student model operates on its own without getting supports from the teacher models during decoding, which makes it eighty times faster and lighter than other typical ensemble methods. All models are evaluated on seven document classification datasets and show a significant advantage over the teacher models for most cases. Our analysis depicts insightful transformation of word embeddings from distillation and suggests a future direction to ensemble approaches using neural models.

Convolutional Neural Networks (CNNs) can achieve excellent computer-assisted diagnosis performance, relying on sufficient annotated training data. Unfortunately, most medical imaging datasets, often collected from various scanners, are small and fragmented. In this context, as a Data Augmentation (DA) technique, Generative Adversarial Networks (GANs) can synthesize realistic/diverse additional training images to fill the data lack in the real image distribution; researchers have improved classification by augmenting images with noise-to-image (e.g., random noise samples to diverse pathological images) or image-to-image GANs (e.g., a benign image to a malignant one). Yet, no research has reported results combining (i) noise-to-image GANs and image-to-image GANs or (ii) GANs and other deep generative models, for further performance boost. Therefore, to maximize the DA effect with the GAN combinations, we propose a two-step GAN-based DA that generates and refines brain MR images with/without tumors separately: (i) Progressive Growing of GANs (PGGANs), multi-stage noise-to-image GAN for high-resolution image generation, first generates realistic/diverse 256 x 256 images--even a physician cannot accurately distinguish them from real ones via Visual Turing Test; (ii) UNsupervised Image-to-image Translation or SimGAN, image-to-image GAN combining GANs/Variational AutoEncoders or using a GAN loss for DA, further refines the texture/shape of the PGGAN-generated images similarly to the real ones. We thoroughly investigate CNN-based tumor classification results, also considering the influence of pre-training on ImageNet and discarding weird-looking GAN-generated images. The results show that, when combined with classic DA, our two-step GAN-based DA can significantly outperform the classic DA alone, in tumor detection (i.e., boosting sensitivity from 93.63% to 97.53%) and also in other tasks.

Using powerful posterior distributions is a popular approach to achieving better variational inference. However, recent works showed that the aggregated posterior may fail to match unit Gaussian prior, thus learning the prior becomes an alternative way to improve the lower-bound. In this paper, for the first time in the literature, we prove the necessity and effectiveness of learning the prior when aggregated posterior does not match unit Gaussian prior, analyze why this situation may happen, and propose a hypothesis that learning the prior may improve reconstruction loss, all of which are supported by our extensive experiment results. We show that using learned Real NVP prior and just one latent variable in VAE, we can achieve test NLL comparable to very deep state-of-the-art hierarchical VAE, outperforming many previous works with complex hierarchical VAE architectures.

We study gradient compression methods to alleviate the communication bottleneck in data-parallel distributed optimization. Despite the significant attention received, current compression schemes either do not scale well, or fail to achieve the target test accuracy. We propose a new low-rank gradient compressor based on power iteration that can i) compress gradients rapidly, ii) efficiently aggregate the compressed gradients using all-reduce, and iii) achieve test performance on par with SGD. The proposed algorithm is the only method evaluated that achieves consistent wall-clock speedups when benchmarked against regular SGD with an optimized communication backend. We demonstrate reduced training times for convolutional networks as well as LSTMs on common datasets. Our code is available at https://github.com/epfml/powersgd.

Graphs are ideal representations of complex, relational information. Their applications span diverse areas of science and engineering, such as Feynman diagrams in fundamental physics, the structures of molecules in chemistry or transport systems in urban planning. Recently, many of these examples turned into the spotlight as applications of machine learning (ML). There, common challenges to the successful deployment of ML are domain-specific constraints, which lead to semantically constrained graphs. While much progress has been achieved in the generation of valid graphs for domain- and model-specific applications, a general approach has not been demonstrated yet. Here, we present a general-purpose, sequence-based, robust representation of semantically constrained graphs, which we call SELFIES (SELF-referencIng Embedded Strings). SELFIES are based on a Chomsky type-2 grammar, augmented with two self-referencing functions. We demonstrate their applicability to represent chemical compound structures and compare them to perhaps the most popular 2D representation, SMILES, and other important baselines. We find stronger robustness against character mutations while still maintaining similar chemical properties. Even entirely random SELFIES produce semantically valid graphs in most of the cases. As feature representation in variational autoencoders, SELFIES provide a substantial improvement in the task of in reconstruction, validity, and diversity. We anticipate that SELFIES allow for direct applications in ML, without the need for domain-specific adaptation of model architectures. SELFIES are not limited to the structures of small molecules, and we show how to apply them to two other examples from the sciences: representations of DNA and interaction graphs for quantum mechanical experiments.

Although optimization is the longstanding algorithmic backbone of machine learning, new models still require the time-consuming implementation of new solvers. As a result, there are thousands of implementations of optimization algorithms for machine learning problems. A natural question is, if it is always necessary to implement a new solver, or if there is one algorithm that is sufficient for most models. Common belief suggests that such a one-algorithm-fits-all approach cannot work, because this algorithm cannot exploit model specific structure and thus cannot be efficient and robust on a wide variety of problems. Here, we challenge this common belief. We have designed and implemented the optimization framework GENO (GENeric Optimization) that combines a modeling language with a generic solver. GENO generates a solver from the declarative specification of an optimization problem class. The framework is flexible enough to encompass most of the classical machine learning problems. We show on a wide variety of classical but also some recently suggested problems that the automatically generated solvers are (1) as efficient as well-engineered specialized solvers, (2) more efficient by a decent margin than recent state-of-the-art solvers, and (3) orders of magnitude more efficient than classical modeling language plus solver approaches.

Searching new molecules in areas like drug discovery often starts from the core structures of candidate molecules to optimize the properties of interest. The way as such has called for a strategy of designing molecules retaining a particular scaffold as a substructure. On this account, our present work proposes a scaffold-based molecular generative model. The model generates molecular graphs by extending the graph of a scaffold through sequential additions of vertices and edges. In contrast to previous related models, our model guarantees the generated molecules to retain the given scaffold with certainty. Our evaluation of the model using unseen scaffolds showed the validity, uniqueness, and novelty of generated molecules as high as the case using seen scaffolds. This confirms that the model can generalize the learned chemical rules of adding atoms and bonds rather than simply memorizing the mapping from scaffolds to molecules during learning. Furthermore, despite the restraint of fixing core structures, our model could simultaneously control multiple molecular properties when generating new molecules.

In this paper, we present a novel attack against authorship attribution of source code. We exploit that recent attribution methods rest on machine learning and thus can be deceived by adversarial examples of source code. Our attack performs a series of semantics-preserving code transformations that mislead learning-based attribution but appear plausible to a developer. The attack is guided by Monte-Carlo tree search that enables us to operate in the discrete domain of source code. In an empirical evaluation with source code from 204 programmers, we demonstrate that our attack has a substantial effect on two recent attribution methods, whose accuracy drops from over 88% to 1% under attack. Furthermore, we show that our attack can imitate the coding style of developers with high accuracy and thereby induce false attributions. We conclude that current approaches for authorship attribution are inappropriate for practical application and there is a need for resilient analysis techniques.

Various encoder-decoder models have been applied to response generation in open-domain dialogs, but a majority of conventional models directly learn a mapping from lexical input to lexical output without explicitly modeling intermediate representations. Utilizing language hierarchy and modeling intermediate information have been shown to benefit many language understanding and generation tasks. Motivated by Broca's aphasia, we propose to use a content word sequence as an intermediate representation for open-domain response generation. Experimental results show that the proposed method improves content relatedness of produced responses, and our models can often choose correct grammar for generated content words. Meanwhile, instead of evaluating complete sentences, we propose to compute conventional metrics on content word sequences, which is a better indicator of content relevance.

We demonstrate the need and potential of systematically integrated vision and semantics} solutions for visual sensemaking (in the backdrop of autonomous driving). A general method for online visual sensemaking using answer set programming is systematically formalised and fully implemented. The method integrates state of the art in (deep learning based) visual computing, and is developed as a modular framework usable within hybrid architectures for perception & control. We evaluate and demo with community established benchmarks KITTIMOD and MOT. As use-case, we focus on the significance of human-centred visual sensemaking ---e.g., semantic representation and explainability, question-answering, commonsense interpolation--- in safety-critical autonomous driving situations.