Professionals as well as the general public need effective assistance to access, understand and consume complex biomedical concepts. For example, doctors always want to be aware of up-to-date clinical evidence for the diagnosis and treatment of diseases under the scheme of Evidence-based Medicine [165], and the general public is becoming increasingly interested in learning about their own health conditions on the Internet [54]. Traditionally, Information Retrieval (IR) systems, such as PubMed, have been used to meet such information needs. However, classical IR is still not efficient enough [71, 77, 99, 164]. For instance, Russell-Rose and Chamberlain [164] reported that it requires 4 expert hours to answer complex medical queries using search engines. Compared with the retrieval systems that typically return a list of relevant documents for the users to read, Question Answering (QA) systems that provide direct answers to users' questions are more straightforward and intuitive. In general, QA itself is a challenging benchmark Natural Language Processing (NLP) task for evaluating the abilities of intelligent systems to understand a question, retrieve and utilize relevant materials and generate its answer. With the rapid development of computing hardware, modern QA models, especially those based on deep learning [30, 31, 42, 146, 171], achieve comparable or even better performance than human on many benchmark datasets [67, 83, 154, 155, 215] and have been successfully adopted in general domain search engines and conversational assistants [150, 236]. The Text REtrieval Conference (TREC) QA Track has triggered the modern QA research [197], when QA models were mostly based on IR.

We propose a benchmark to measure whether a language model is truthful in generating answers to questions. The benchmark comprises 817 questions that span 38 categories, including health, law, finance and politics. We crafted questions that some humans would answer falsely due to a false belief or misconception. To perform well, models must avoid generating false answers learned from imitating human texts. We tested GPT-3, GPT-Neo/J, GPT-2 and a T5-based model. The best model was truthful on 58% of questions, while human performance was 94%. Models generated many false answers that mimic popular misconceptions and have the potential to deceive humans. The largest models were generally the least truthful. For example, the 6B-parameter GPT-J model was 17% less truthful than its 125M-parameter counterpart. This contrasts with other NLP tasks, where performance improves with model size. However, this result is expected if false answers are learned from the training distribution. We suggest that scaling up models alone is less promising for improving truthfulness than fine-tuning using training objectives other than imitation of text from the web.

The Automatic Dependent Surveillance Broadcast protocol is one of the latest compulsory advances in air surveillance. While it supports the tracking of the ever-growing number of aircraft in the air, it also introduces cybersecurity issues that must be mitigated e.g., false data injection attacks where an attacker emits fake surveillance information. The recent data sources and tools available to obtain flight tracking records allow the researchers to create datasets and develop Machine Learning models capable of detecting such anomalies in En-Route trajectories. In this context, we propose a novel multivariate anomaly detection model called Discriminatory Auto-Encoder (DAE). It uses the baseline of a regular LSTM-based auto-encoder but with several decoders, each getting data of a specific flight phase (e.g. climbing, cruising or descending) during its training.To illustrate the DAE's efficiency, an evaluation dataset was created using real-life anomalies as well as realistically crafted ones, with which the DAE as well as three anomaly detection models from the literature were evaluated. Results show that the DAE achieves better results in both accuracy and speed of detection. The dataset, the models implementations and the evaluation results are available in an online repository, thereby enabling replicability and facilitating future experiments.

Abstract--Modelling trajectory in general, and vessel trajectory in particular, is a difficult task because of the multimodal and complex nature of motion data. In this paper, we present TrAISformer--a novel deep learning architecture that can forecast vessel positions using AIS (Automatic Identification System) observations. We address the multimodality by introducing a discrete representation of AIS data and re-frame the prediction, which is originally a regression problem, as a classification problem. The model encodes complex movement patterns in AIS data in high-dimensional vectors, then applies a transformer to extract useful long-term correlations from sequences of those embeddings to sample future vessel positions. Experimental results on real, public AIS data demonstrate that TrAISformer significantly outperforms state-of-the-art methods.

While recent research on natural language inference has considerably benefited from large annotated datasets, the amount of inference-related knowledge (including commonsense) provided in the annotated data is still rather limited. There have been two lines of approaches that can be used to further address the limitation: (1) unsupervised pretraining can leverage knowledge in much larger unstructured text data; (2) structured (often human-curated) knowledge has started to be considered in neural-network-based models for NLI. An immediate question is whether these two approaches complement each other, or how to develop models that can bring together their advantages. In this paper, we propose models that leverage structured knowledge in different components of pre-trained models. Our results show that the proposed models perform better than previous BERT-based state-of-the-art models. Although our models are proposed for NLI, they can be easily extended to other sentence or sentence-pair classification problems.

We demonstrate that, hidden within one-layer randomly weighted neural networks, there exist subnetworks that can achieve impressive performance, without ever modifying the weight initializations, on machine translation tasks. To find subnetworks for one-layer randomly weighted neural networks, we apply different binary masks to the same weight matrix to generate different layers. Hidden within a one-layer randomly weighted Transformer, we find that subnetworks that can achieve 29.45/17.29 BLEU on IWSLT14/WMT14. Using a fixed pre-trained embedding layer, the previously found subnetworks are smaller than, but can match 98%/92% (34.14/25.24 BLEU) of the performance of, a trained Transformer small/base on IWSLT14/WMT14. Furthermore, we demonstrate the effectiveness of larger and deeper transformers in this setting, as well as the impact of different initialization methods. We released the source code at https://github.com/sIncerass/one_layer_lottery_ticket.

We investigate the use of multimodal information contained in images as an effective method for enhancing the commonsense of Transformer models for text generation. We perform experiments using BART and T5 on concept-to-text generation, specifically the task of generative commonsense reasoning, or CommonGen. We call our approach VisCTG: Visually Grounded Concept-to-Text Generation. VisCTG involves captioning images representing appropriate everyday scenarios, and using these captions to enrich and steer the generation process. Comprehensive evaluation and analysis demonstrate that VisCTG noticeably improves model performance while successfully addressing several issues of the baseline generations, including poor commonsense, fluency, and specificity.

Neural network controllers have become popular in control tasks thanks to their flexibility and expressivity. Stability is a crucial property for safety-critical dynamical systems, while stabilization of partially observed systems, in many cases, requires controllers to retain and process long-term memories of the past. We consider the important class of recurrent neural networks (RNN) as dynamic controllers for nonlinear uncertain partially-observed systems, and derive convex stability conditions based on integral quadratic constraints, S-lemma and sequential convexification. To ensure stability during the learning and control process, we propose a projected policy gradient method that iteratively enforces the stability conditions in the reparametrized space taking advantage of mild additional information on system dynamics. Numerical experiments show that our method learns stabilizing controllers while using fewer samples and achieving higher final performance compared with policy gradient.

Mixture-of-Experts (MoE) with sparse conditional computation has been proved an effective architecture for scaling attention-based models to more parameters with comparable computation cost. In this paper, we propose Sparse-MLP, scaling the recent MLP-Mixer model with sparse MoE layers, to achieve a more computation-efficient architecture. We replace a subset of dense MLP blocks in the MLP-Mixer model with Sparse blocks. In each Sparse block, we apply two stages of MoE layers: one with MLP experts mixing information within channels along image patch dimension, one with MLP experts mixing information within patches along the channel dimension. Besides, to reduce computational cost in routing and improve expert capacity, we design Re-represent layers in each Sparse block. These layers are to re-scale image representations by two simple but effective linear transformations. When pre-training on ImageNet-1k with MoCo v3 algorithm, our models can outperform dense MLP models by 2.5\% on ImageNet Top-1 accuracy with fewer parameters and computational cost. On small-scale downstream image classification tasks, i.e. Cifar10 and Cifar100, our Sparse-MLP can still achieve better performance than baselines.

With the constant increase of the number of quantum bits (qubits) in the actual quantum computers, implementing and accelerating the prevalent deep learning on quantum computers are becoming possible. Along with this trend, there emerge quantum neural architectures based on different designs of quantum neurons. A fundamental question in quantum deep learning arises: what is the best quantum neural architecture? Inspired by the design of neural architectures for classical computing which typically employs multiple types of neurons, this paper makes the very first attempt to mix quantum neuron designs to build quantum neural architectures. We observe that the existing quantum neuron designs may be quite different but complementary, such as neurons from variation quantum circuits (VQC) and Quantumflow. More specifically, VQC can apply real-valued weights but suffer from being extended to multiple layers, while QuantumFlow can build a multi-layer network efficiently, but is limited to use binary weights. To take their respective advantages, we propose to mix them together and figure out a way to connect them seamlessly without additional costly measurement. We further investigate the design principles to mix quantum neurons, which can provide guidance for quantum neural architecture exploration in the future. Experimental results demonstrate that the identified quantum neural architectures with mixed quantum neurons can achieve 90.62% of accuracy on the MNIST dataset, compared with 52.77% and 69.92% on the VQC and QuantumFlow, respectively.