Dirk Weissenborn

Matthias Minderer, Alexey Gritsenko, Austin Stone et al.

Combining simple architectures with large-scale pre-training has led to massive improvements in image classification. For object detection, pre-training and scaling approaches are less well established, especially in the long-tailed and open-vocabulary setting, where training data is relatively scarce. In this paper, we propose a strong recipe for transferring image-text models to open-vocabulary object detection. We use a standard Vision Transformer architecture with minimal modifications, contrastive image-text pre-training, and end-to-end detection fine-tuning. Our analysis of the scaling properties of this setup shows that increasing image-level pre-training and model size yield consistent improvements on the downstream detection task. We provide the adaptation strategies and regularizations needed to attain very strong performance on zero-shot text-conditioned and one-shot image-conditioned object detection. Code and models are available on GitHub.

Manoj Kumar, Dirk Weissenborn, Nal Kalchbrenner

We present the Colorization Transformer, a novel approach for diverse high fidelity image colorization based on self-attention. Given a grayscale image, the colorization proceeds in three steps. We first use a conditional autoregressive transformer to produce a low resolution coarse coloring of the grayscale image. Our architecture adopts conditional transformer layers to effectively condition grayscale input. Two subsequent fully parallel networks upsample the coarse colored low resolution image into a finely colored high resolution image. Sampling from the Colorization Transformer produces diverse colorings whose fidelity outperforms the previous state-of-the-art on colorising ImageNet based on FID results and based on a human evaluation in a Mechanical Turk test. Remarkably, in more than 60% of cases human evaluators prefer the highest rated among three generated colorings over the ground truth. The code and pre-trained checkpoints for Colorization Transformer are publicly available at https://github.com/google-research/google-research/tree/master/coltran

Jonathan Ho, Nal Kalchbrenner, Dirk Weissenborn et al.

We propose Axial Transformers, a self-attention-based autoregressive model for images and other data organized as high dimensional tensors. Existing autoregressive models either suffer from excessively large computational resource requirements for high dimensional data, or make compromises in terms of distribution expressiveness or ease of implementation in order to decrease resource requirements. Our architecture, by contrast, maintains both full expressiveness over joint distributions over data and ease of implementation with standard deep learning frameworks, while requiring reasonable memory and computation and achieving state-of-the-art results on standard generative modeling benchmarks. Our models are based on axial attention, a simple generalization of self-attention that naturally aligns with the multiple dimensions of the tensors in both the encoding and the decoding settings. Notably the proposed structure of the layers allows for the vast majority of the context to be computed in parallel during decoding without introducing any independence assumptions. This semi-parallel structure goes a long way to making decoding from even a very large Axial Transformer broadly applicable. We demonstrate state-of-the-art results for the Axial Transformer on the ImageNet-32 and ImageNet-64 image benchmarks as well as on the BAIR Robotic Pushing video benchmark. We open source the implementation of Axial Transformers.

Jean-Baptiste Cordonnier, Aravindh Mahendran, Alexey Dosovitskiy et al.

Neural Networks require large amounts of memory and compute to process high resolution images, even when only a small part of the image is actually informative for the task at hand. We propose a method based on a differentiable Top-K operator to select the most relevant parts of the input to efficiently process high resolution images. Our method may be interfaced with any downstream neural network, is able to aggregate information from different patches in a flexible way, and allows the whole model to be trained end-to-end using backpropagation. We show results for traffic sign recognition, inter-patch relationship reasoning, and fine-grained recognition without using object/part bounding box annotations during training.

Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov et al.

While the Transformer architecture has become the de-facto standard for natural language processing tasks, its applications to computer vision remain limited. In vision, attention is either applied in conjunction with convolutional networks, or used to replace certain components of convolutional networks while keeping their overall structure in place. We show that this reliance on CNNs is not necessary and a pure transformer applied directly to sequences of image patches can perform very well on image classification tasks. When pre-trained on large amounts of data and transferred to multiple mid-sized or small image recognition benchmarks (ImageNet, CIFAR-100, VTAB, etc.), Vision Transformer (ViT) attains excellent results compared to state-of-the-art convolutional networks while requiring substantially fewer computational resources to train.

Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner et al.

Learning object-centric representations of complex scenes is a promising step towards enabling efficient abstract reasoning from low-level perceptual features. Yet, most deep learning approaches learn distributed representations that do not capture the compositional properties of natural scenes. In this paper, we present the Slot Attention module, an architectural component that interfaces with perceptual representations such as the output of a convolutional neural network and produces a set of task-dependent abstract representations which we call slots. These slots are exchangeable and can bind to any object in the input by specializing through a competitive procedure over multiple rounds of attention. We empirically demonstrate that Slot Attention can extract object-centric representations that enable generalization to unseen compositions when trained on unsupervised object discovery and supervised property prediction tasks.

Dirk Weissenborn, Oscar Täckström, Jakob Uszkoreit

Due to the statistical complexity of video, the high degree of inherent stochasticity, and the sheer amount of data, generating natural video remains a challenging task. State-of-the-art video generation models often attempt to address these issues by combining sometimes complex, usually video-specific neural network architectures, latent variable models, adversarial training and a range of other methods. Despite their often high complexity, these approaches still fall short of generating high quality video continuations outside of narrow domains and often struggle with fidelity. In contrast, we show that conceptually simple autoregressive video generation models based on a three-dimensional self-attention mechanism achieve competitive results across multiple metrics on popular benchmark datasets, for which they produce continuations of high fidelity and realism. We also present results from training our models on Kinetics, a large scale action recognition dataset comprised of YouTube videos exhibiting phenomena such as camera movement, complex object interactions and diverse human movement. While modeling these phenomena consistently remains elusive, we hope that our results, which include occasional realistic continuations encourage further research on comparatively complex, large scale datasets such as Kinetics.

Dirk Weissenborn, Pasquale Minervini, Tim Dettmers et al.

Many Machine Reading and Natural Language Understanding tasks require reading supporting text in order to answer questions. For example, in Question Answering, the supporting text can be newswire or Wikipedia articles; in Natural Language Inference, premises can be seen as the supporting text and hypotheses as questions. Providing a set of useful primitives operating in a single framework of related tasks would allow for expressive modelling, and easier model comparison and replication. To that end, we present Jack the Reader (Jack), a framework for Machine Reading that allows for quick model prototyping by component reuse, evaluation of new models on existing datasets as well as integrating new datasets and applying them on a growing set of implemented baseline models. Jack is currently supporting (but not limited to) three tasks: Question Answering, Natural Language Inference, and Link Prediction. It is developed with the aim of increasing research efficiency and code reuse.

Roland Roller, Madeleine Kittner, Dirk Weissenborn et al.

Biomedical concept normalization links concept mentions in texts to a semantically equivalent concept in a biomedical knowledge base. This task is challenging as concepts can have different expressions in natural languages, e.g. paraphrases, which are not necessarily all present in the knowledge base. Concept normalization of non-English biomedical text is even more challenging as non-English resources tend to be much smaller and contain less synonyms. To overcome the limitations of non-English terminologies we propose a cross-lingual candidate search for concept normalization using a character-based neural translation model trained on a multilingual biomedical terminology. Our model is trained with Spanish, French, Dutch and German versions of UMLS. The evaluation of our model is carried out on the French Quaero corpus, showing that it outperforms most teams of CLEF eHealth 2015 and 2016. Additionally, we compare performance to commercial translators on Spanish, French, Dutch and German versions of Mantra. Our model performs similarly well, but is free of charge and can be run locally. This is particularly important for clinical NLP applications as medical documents underlay strict privacy restrictions.

Georg Wiese, Dirk Weissenborn, Mariana Neves

This paper describes our submission to the 2017 BioASQ challenge. We participated in Task B, Phase B which is concerned with biomedical question answering (QA). We focus on factoid and list question, using an extractive QA model, that is, we restrict our system to output substrings of the provided text snippets. At the core of our system, we use FastQA, a state-of-the-art neural QA system. We extended it with biomedical word embeddings and changed its answer layer to be able to answer list questions in addition to factoid questions. We pre-trained the model on a large-scale open-domain QA dataset, SQuAD, and then fine-tuned the parameters on the BioASQ training set. With our approach, we achieve state-of-the-art results on factoid questions and competitive results on list questions.

Georg Wiese, Dirk Weissenborn, Mariana Neves

Factoid question answering (QA) has recently benefited from the development of deep learning (DL) systems. Neural network models outperform traditional approaches in domains where large datasets exist, such as SQuAD (ca. 100,000 questions) for Wikipedia articles. However, these systems have not yet been applied to QA in more specific domains, such as biomedicine, because datasets are generally too small to train a DL system from scratch. For example, the BioASQ dataset for biomedical QA comprises less then 900 factoid (single answer) and list (multiple answers) QA instances. In this work, we adapt a neural QA system trained on a large open-domain dataset (SQuAD, source) to a biomedical dataset (BioASQ, target) by employing various transfer learning techniques. Our network architecture is based on a state-of-the-art QA system, extended with biomedical word embeddings and a novel mechanism to answer list questions. In contrast to existing biomedical QA systems, our system does not rely on domain-specific ontologies, parsers or entity taggers, which are expensive to create. Despite this fact, our systems achieve state-of-the-art results on factoid questions and competitive results on list questions.

Dirk Weissenborn, Tomáš Kočiský, Chris Dyer

Common-sense and background knowledge is required to understand natural language, but in most neural natural language understanding (NLU) systems, this knowledge must be acquired from training corpora during learning, and then it is static at test time. We introduce a new architecture for the dynamic integration of explicit background knowledge in NLU models. A general-purpose reading module reads background knowledge in the form of free-text statements (together with task-specific text inputs) and yields refined word representations to a task-specific NLU architecture that reprocesses the task inputs with these representations. Experiments on document question answering (DQA) and recognizing textual entailment (RTE) demonstrate the effectiveness and flexibility of the approach. Analysis shows that our model learns to exploit knowledge in a semantically appropriate way.

Dirk Weissenborn, Georg Wiese, Laura Seiffe

Recent development of large-scale question answering (QA) datasets triggered a substantial amount of research into end-to-end neural architectures for QA. Increasingly complex systems have been conceived without comparison to simpler neural baseline systems that would justify their complexity. In this work, we propose a simple heuristic that guides the development of neural baseline systems for the extractive QA task. We find that there are two ingredients necessary for building a high-performing neural QA system: first, the awareness of question words while processing the context and second, a composition function that goes beyond simple bag-of-words modeling, such as recurrent neural networks. Our results show that FastQA, a system that meets these two requirements, can achieve very competitive performance compared with existing models. We argue that this surprising finding puts results of previous systems and the complexity of recent QA datasets into perspective.

Shinichi Nakajima, Sebastian Krause, Dirk Weissenborn et al.

In relation extraction, a key process is to obtain good detectors that find relevant sentences describing the target relation. To minimize the necessity of labeled data for refining detectors, previous work successfully made use of BabelNet, a semantic graph structure expressing relationships between synsets, as side information or prior knowledge. The goal of this paper is to enhance the use of graph structure in the framework of random walk with a few adjustable parameters. Actually, a straightforward application of random walk degrades the performance even after parameter optimization. With the insight from this unsuccessful trial, we propose SynsetRank, which adjusts the initial probability so that high degree nodes influence the neighbors as strong as low degree nodes. In our experiment on 13 relations in the FB15K-237 dataset, SynsetRank significantly outperforms baselines and the plain random walk approach.

Dirk Weissenborn

We present a novel neural architecture for answering queries, designed to optimally leverage explicit support in the form of query-answer memories. Our model is able to refine and update a given query while separately accumulating evidence for predicting the answer. Its architecture reflects this separation with dedicated embedding matrices and loosely connected information pathways (modules) for updating the query and accumulating evidence. This separation of responsibilities effectively decouples the search for query related support and the prediction of the answer. On recent benchmark datasets for reading comprehension, our model achieves state-of-the-art results. A qualitative analysis reveals that the model effectively accumulates weighted evidence from the query and over multiple support retrieval cycles which results in a robust answer prediction.

Dirk Weissenborn

Many important NLP problems can be posed as dual-sequence or sequence-to-sequence modeling tasks. Recent advances in building end-to-end neural architectures have been highly successful in solving such tasks. In this work we propose a new architecture for dual-sequence modeling that is based on associative memory. We derive AM-RNNs, a recurrent associative memory (AM) which augments generic recurrent neural networks (RNN). This architecture is extended to the Dual AM-RNN which operates on two AMs at once. Our models achieve very competitive results on textual entailment. A qualitative analysis demonstrates that long range dependencies between source and target-sequence can be bridged effectively using Dual AM-RNNs. However, an initial experiment on auto-encoding reveals that these benefits are not exploited by the system when learning to solve sequence-to-sequence tasks which indicates that additional supervision or regularization is needed.

Dirk Weissenborn, Tim Rocktäschel

Recurrent neural networks such as the GRU and LSTM found wide adoption in natural language processing and achieve state-of-the-art results for many tasks. These models are characterized by a memory state that can be written to and read from by applying gated composition operations to the current input and the previous state. However, they only cover a small subset of potentially useful compositions. We propose Multi-Function Recurrent Units (MuFuRUs) that allow for arbitrary differentiable functions as composition operations. Furthermore, MuFuRUs allow for an input- and state-dependent choice of these composition operations that is learned. Our experiments demonstrate that the additional functionality helps in different sequence modeling tasks, including the evaluation of propositional logic formulae, language modeling and sentiment analysis.