While early research in neural architecture search (NAS) required extreme computational resources, the recent releases of tabular and surrogate benchmarks have greatly increased the speed and reproducibility of NAS research. However, two of the most popular benchmarks do not provide the full training information for each architecture. As a result, on these benchmarks it is not possible to run many types of multi-fidelity techniques, such as learning curve extrapolation, that require evaluating architectures at arbitrary epochs. In this work, we present a method using singular value decomposition and noise modeling to create surrogate benchmarks, NAS-Bench-111, NAS-Bench-311, and NAS-Bench-NLP11, that output the full training information for each architecture, rather than just the final validation accuracy. We demonstrate the power of using the full training information by introducing a learning curve extrapolation framework to modify single-fidelity algorithms, showing that it leads to improvements over popular single-fidelity algorithms which claimed to be state-of-the-art upon release.

In this paper, we evaluate the quality of knowledge representations encoded in deep neural networks (DNNs) for 3D point cloud processing. We propose a method to disentangle the overall model vulnerability into the sensitivity to the rotation, the translation, the scale, and local 3D structures. Besides, we also propose metrics to evaluate the spatial smoothness of encoding 3D structures, and the representation complexity of the DNN. Based on such analysis, experiments expose representation problems with classic DNNs, and explain the utility of the adversarial training.

Automatic detection of click-bait and incongruent news headlines is crucial to maintaining the reliability of the Web and has raised much research attention. However, most existing methods perform poorly when news headlines contain contextually important cardinal values, such as a quantity or an amount. In this work, we focus on this particular case and propose a neural attention based solution, which uses a novel cardinal Part of Speech (POS) tag pattern based hierarchical attention network, namely POSHAN, to learn effective representations of sentences in a news article. In addition, we investigate a novel cardinal phrase guided attention, which uses word embeddings of the contextually-important cardinal value and neighbouring words. In the experiments conducted on two publicly available datasets, we observe that the proposed methodgives appropriate significance to cardinal values and outperforms all the baselines. An ablation study of POSHAN shows that the cardinal POS-tag pattern-based hierarchical attention is very effective for the cases in which headlines contain cardinal values.

Algorithmic design in neural architecture search (NAS) has received a lot of attention, aiming to improve performance and reduce computational cost. Despite the great advances made, few authors have proposed to tailor initialization techniques for NAS. However, literature shows that a good initial set of solutions facilitate finding the optima. Therefore, in this study, we propose a data-driven technique to initialize a population-based NAS algorithm. Particularly, we proposed a two-step methodology. First, we perform a calibrated clustering analysis of the search space, and second, we extract the centroids and use them to initialize a NAS algorithm. We benchmark our proposed approach against random and Latin hypercube sampling initialization using three population-based algorithms, namely a genetic algorithm, evolutionary algorithm, and aging evolution, on CIFAR-10. More specifically, we use NAS-Bench-101 to leverage the availability of NAS benchmarks. The results show that compared to random and Latin hypercube sampling, the proposed initialization technique enables achieving significant long-term improvements for two of the search baselines, and sometimes in various search scenarios (various training budgets). Moreover, we analyze the distributions of solutions obtained and find that that the population provided by the data-driven initialization technique enables retrieving local optima (maxima) of high fitness and similar configurations.

This paper proposes a method to visualize the discrimination power of intermediate-layer visual patterns encoded by a DNN. Specifically, we visualize (1) how the DNN gradually learns regional visual patterns in each intermediate layer during the training process, and (2) the effects of the DNN using non-discriminative patterns in low layers to construct disciminative patterns in middle/high layers through the forward propagation. Based on our visualization method, we can quantify knowledge points (i.e., the number of discriminative visual patterns) learned by the DNN to evaluate the representation capacity of the DNN. Furthermore, this method also provides new insights into signal-processing behaviors of existing deep-learning techniques, such as adversarial attacks and knowledge distillation.

Automated vehicles are a fundamental component of intelligent transportation systems and will improve safety, traffic efficiency and driving experience, and reduce human errors [1]. Deep learning solutions are used in several autonomous vehicle subsystems in order to perform perception, sensor fusion, scene analysis, and path planning (e.g., [2, 3, 4, 5, 6]). State-of-the-art and human-competitive performance have been achieved by such algorithms on many computer vision tasks related to autonomous vehicles [7]. Nevertheless, over the last years it was demonstrated that deep-learning-based solutions are susceptible to various threats and vulnerabilities that can cause the autonomous vehicles to misbehave in unexpected and potentially dangerous ways. Besides physical attacks that can induce erroneous behaviour there is also the possibility that the camera data can be manipulated directly thus eliciting false algorithmic inferences. This can cause an AI-based perception module/controller to make incorrect decisions, such as when an autonomous vehicle fails to detect a lane/ parking marking and results in a collision [8, 9]. An example of such an attack can be either data injection by malicious software installation [10] in the CAN bus, projecting patterns [11], manipulation of over the air updates [12, 13], or even faults [14]. Hence crafting appropriate defenses against attacks that go beyond the hardware layer is important to realize safe autonomous driving [9, 15, 8].

In this technical report, we present our solution to the Traffic4Cast 2021 Core Challenge, in which participants were asked to develop algorithms for predicting a traffic state 60 minutes ahead, based on the information from the previous hour, in 4 different cities. In contrast to the previously held competitions, this year's challenge focuses on the temporal domain shift in traffic due to the COVID-19 pandemic. Following the past success of U-Net, we utilize it for predicting future traffic maps. Additionally, we explore the usage of pre-trained encoders such as DenseNet and EfficientNet and employ multiple domain adaptation techniques to fight the domain shift. Our solution has ranked third in the final competition.

State-of-the-art approaches for metaphor detection compare their literal - or core - meaning and their contextual meaning using sequential metaphor classifiers based on neural networks. The signal that represents the literal meaning is often represented by (non-contextual) word embeddings. However, metaphorical expressions evolve over time due to various reasons, such as cultural and societal impact. Metaphorical expressions are known to co-evolve with language and literal word meanings, and even drive, to some extent, this evolution. This rises the question whether different, possibly time-specific, representations of literal meanings may impact on the metaphor detection task. To the best of our knowledge, this is the first study which examines the metaphor detection task with a detailed exploratory analysis where different temporal and static word embeddings are used to account for different representations of literal meanings. Our experimental analysis is based on three popular benchmarks used for metaphor detection and word embeddings extracted from different corpora and temporally aligned to different state-of-the-art approaches. The results suggest that different word embeddings do impact on the metaphor detection task and some temporal word embeddings slightly outperform static methods on some performance measures. However, results also suggest that temporal word embeddings may provide representations of words' core meaning even too close to their metaphorical meaning, thus confusing the classifier. Overall, the interaction between temporal language evolution and metaphor detection appears tiny in the benchmark datasets used in our experiments. This suggests that future work for the computational analysis of this important linguistic phenomenon should first start by creating a new dataset where this interaction is better represented.

Dialogue summarization has been extensively studied and applied, where the prior works mainly focused on exploring superior model structures to align the input dialogue and the output summary. However, for professional dialogues (e.g., legal debate and medical diagnosis), semantic/statistical alignment can hardly fill the logical/factual gap between input dialogue discourse and summary output with external knowledge. In this paper, we mainly investigate the factual inconsistency problem for Dialogue Inspectional Summarization (DIS) under non-pretraining and pretraining settings. An innovative end-to-end dialogue summary generation framework is proposed with two auxiliary tasks: Expectant Factual Aspect Regularization (EFAR) and Missing Factual Entity Discrimination (MFED). Comprehensive experiments demonstrate that the proposed model can generate a more readable summary with accurate coverage of factual aspects as well as informing the user with potential missing facts detected from the input dialogue for further human intervention.

Unsupervised graph representation learning is a non-trivial topic for graph data. The success of contrastive learning and self-supervised learning in the unsupervised representation learning of structured data inspires similar attempts on the graph. The current unsupervised graph representation learning and pre-training using the contrastive loss are mainly based on the contrast between handcrafted augmented graph data. However, the graph data augmentation is still not well-explored due to the unpredictable invariance. In this paper, we propose a novel collaborative graph neural networks contrastive learning framework (CGCL), which uses multiple graph encoders to observe the graph. Features observed from different views act as the graph augmentation for contrastive learning between graph encoders, avoiding any perturbation to guarantee the invariance. CGCL is capable of handling both graph-level and node-level representation learning. Extensive experiments demonstrate the advantages of CGCL in unsupervised graph representation learning and the non-necessity of handcrafted data augmentation composition for graph representation learning.