Learning interpretable dialog structure from human-human dialogs yields basic insights into the structure of conversation, and also provides background knowledge to facilitate dialog generation. In this paper, we conduct unsupervised discovery of dialog structure from chitchat corpora, and then leverage it to facilitate dialog generation in downstream systems. To this end, we present a Discrete Variational Auto-Encoder with Graph Neural Network (DVAE-GNN), to discover a unified human-readable dialog structure. The structure is a two-layer directed graph that contains session-level semantics in the upper-layer vertices, utterance-level semantics in the lower-layer vertices, and edges among these semantic vertices. In particular, we integrate GNN into DVAE to fine-tune utterance-level semantics for more effective recognition of session-level semantic vertex. Furthermore, to alleviate the difficulty of discovering a large number of utterance-level semantics, we design a coupling mechanism that binds each utterance-level semantic vertex with a distinct phrase to provide prior semantics. Experimental results on two benchmark corpora confirm that DVAE-GNN can discover meaningful dialog structure, and the use of dialog structure graph as background knowledge can facilitate a graph grounded conversational system to conduct coherent multi-turn dialog generation.

A video-grounded dialogue system is required to understand both dialogue, which contains semantic dependencies from turn to turn, and video, which contains visual cues of spatial and temporal scene variations. Building such dialogue systems is a challenging problem involving complex multimodal and temporal inputs, and studying them independently is hard with existing datasets. Existing benchmarks do not have enough annotations to help analyze dialogue systems and understand their linguistic and visual reasoning capability and limitations in isolation. These benchmarks are also not explicitly designed to minimize biases that models can exploit without actual reasoning. To address these limitations, in this paper, we present a diagnostic dataset that can test a range of reasoning abilities on videos and dialogues. The dataset is designed to contain minimal biases and has detailed annotations for the different types of reasoning each question requires, including cross-turn video interval tracking and dialogue object tracking. We use our dataset to analyze several dialogue system approaches, providing interesting insights into their abilities and limitations. In total, the dataset contains $10$ instances of $10$-round dialogues for each of $\sim11k$ synthetic videos, resulting in more than $100k$ dialogues and $1M$ question-answer pairs. Our code and dataset will be made public.

The ability to capture complex linguistic structures and long-term dependencies among words in the passage is essential for many natural language understanding tasks. In relation extraction, dependency trees that contain rich syntactic clues have been widely used to help capture long-term dependencies in text. Graph neural networks (GNNs), one of the means to encode dependency graphs, has been shown effective in several prior works. However, relatively little attention has been paid to the receptive fields of GNNs, which can be crucial in tasks with extremely long text that go beyond single sentences and require discourse analysis. In this work, we leverage the idea of graph pooling and propose the Mirror Graph Convolution Network (MrGCN), a GNN model with pooling-unpooling structures tailored to relation extraction. The pooling branch reduces the graph size and enables the GCN to obtain larger receptive fields within less layers; the unpooling branch restores the pooled graph to its original resolution such that token-level relation extraction can be performed. Experiments on two datasets demonstrate the effectiveness of our method, showing significant improvements over previous results.

Thumbnail is the face of online videos. The explosive growth of videos both in number and variety underpins the importance of a good thumbnail because it saves potential viewers time to choose videos and even entice them to click on them. A good thumbnail should be a frame that best represents the content of a video while at the same time capturing viewers' attention. However, the techniques and models in the past only focus on frames within a video, and we believe such narrowed focus leave out much useful information that are part of a video. In this paper, we expand the definition of content to include title, description, and audio of a video and utilize information provided by these modalities in our selection model. Specifically, our model will first sample frames uniformly in time and return the top 1,000 frames in this subset with the highest aesthetic scores by a Double-column Convolutional Neural Network, to avoid the computational burden of processing all frames in downstream task. Then, the model incorporates frame features extracted from VGG16, text features from ELECTRA, and audio features from TRILL. These models were selected because of their results on popular datasets as well as their competitive performances. After feature extraction, the time-series features, frames and audio, will be fed into Transformer encoder layers to return a vector representing their corresponding modality. Each of the four features (frames, title, description, audios) will pass through a context gating layer before concatenation. Finally, our model will generate a vector in the latent space and select the frame that is most similar to this vector in the latent space. To the best of our knowledge, we are the first to propose a multi-modal deep learning model to select video thumbnail, which beats the result from the previous State-of-The-Art models.

Deep, heavily overparameterized language models such as BERT, XLNet and T5 have achieved impressive success in many natural language processing (NLP) tasks. However, their high model complexity requires enormous computation resources and extremely long training time for both pre-training and fine-tuning. Many works have studied model compression on large NLP models, but only focusing on reducing inference time while still requiring an expensive training process. Other works use extremely large batch sizes to shorten the pre-training time, at the expense of higher computational resource demands. In this paper, inspired by the Early-Bird Lottery Tickets recently studied for computer vision tasks, we propose EarlyBERT, a general computationally-efficient training algorithm applicable to both pre-training and fine-tuning of large-scale language models. By slimming the self-attention and fully-connected sub-layers inside a transformer, we are the first to identify structured winning tickets in the early stage of BERT training. We apply those tickets towards efficient BERT training, and conduct comprehensive pre-training and fine-tuning experiments on GLUE and SQuAD downstream tasks. Our results show that EarlyBERT achieves comparable performance to standard BERT, with 35 45% less training time. Large-scale pre-trained language models (e.g., BERT (Devlin et al., 2018), XLNet (Yang et al., 2019), T5 (Raffel et al., 2019)) have significantly advanced the state of the art in the NLP field.

Graph Neural Networks (GNNs) have emerged as a powerful and flexible framework for representation learning on irregular data. As they generalize the operations of classical CNNs on grids to arbitrary topologies, GNNs also bring much of the implementation challenges of their Euclidean counterparts. Model size, memory footprint, and energy consumption are common concerns for many real-world applications. Network binarization allocates a single bit to network parameters and activations, thus dramatically reducing the memory requirements (up to 32x compared to single-precision floating-point parameters) and maximizing the benefits of fast SIMD instructions of modern hardware for measurable speedups. However, in spite of the large body of work on binarization for classical CNNs, this area remains largely unexplored in geometric deep learning. In this paper, we present and evaluate different strategies for the binarization of graph neural networks. We show that through careful design of the models, and control of the training process, binary graph neural networks can be trained at only a moderate cost in accuracy on challenging benchmarks. In particular, we present the first dynamic graph neural network in Hamming space, able to leverage efficient $k$-NN search on binary vectors to speed-up the construction of the dynamic graph. We further verify that the binary models offer significant savings on embedded devices.

We apply methods of topological data analysis to loss functions to gain insights on learning of deep neural networks and their generalization properties We study global properties of the loss function's gradient flow. We use topological data analysis of the loss function and its Morse complex to relate local behaviour along gradient trajectories with global properties of the loss surface. We define neural network's Topological Obstructions' score («TOscore») with help of robust topological invariants (barcodes of loss function) that quantify the "badness" of local minima for gradient-based optimization. We have made several experiments for computing these invariants, for small neural networks, and for fully connected, convolutional and ResNetlike neural networks on different datasets: MNIST, Fashion MNIST, CIFAR10, SVHN. Our two principal observations are 1) the neural network's barcode and TOscore decrease with the increase of the neural network's depth and width 2) there is an intriguing connection between the length of minima's segments in the barcode and the minima's generalization error. Introduction Mathematically, if one opens the "black box" of deep learning, there are two immediate mysteries.

This paper addresses the problem of computing dense correspondence between 3D shapes in the form of point clouds, which is a challenging and fundamental problem in computer vision and digital geometry processing. Conventional approaches often solve the problem in a supervised manner, requiring massive annotated data, which is difficult and/or expensive to obtain. Motivated by the intuition that one can transform two aligned point clouds to each other more easily and meaningfully than a misaligned pair, we propose CorrNet3D -- the first unsupervised and end-to-end deep learning-based framework -- to drive the learning of dense correspondence by means of deformation-like reconstruction to overcome the need for annotated data. Specifically, CorrNet3D consists of a deep feature embedding module and two novel modules called correspondence indicator and symmetric deformation. Feeding a pair of raw point clouds, our model first learns the pointwise features and passes them into the indicator to generate a learnable correspondence matrix used to permute the input pair. The symmetric deformer, with an additional regularized loss, transforms the two permuted point clouds to each other to drive the unsupervised learning of the correspondence. The extensive experiments on both synthetic and real-world datasets of rigid and non-rigid 3D shapes show our CorrNet3D outperforms state-of-the-art methods to a large extent, including those taking meshes as input. CorrNet3D is a flexible framework in that it can be easily adapted to supervised learning if annotated data are available.

Memorability determines what evanesces into emptiness, and what worms its way into the deepest furrows of our minds. It is the key to curating more meaningful media content as we wade through daily digital torrents. The Predicting Media Memorability task in MediaEval 2020 aims to address the question of media memorability by setting the task of automatically predicting video memorability. Our approach is a multimodal deep learning-based late fusion that combines visual, semantic, and auditory features. We used audio gestalt to estimate the influence of the audio modality on overall video memorability, and accordingly inform which combination of features would best predict a given video's memorability scores.

Deep generative models have achieved great success in areas such as image, speech, and natural language processing in the past few years. Thanks to the advances in graph-based deep learning, and in particular graph representation learning, deep graph generation methods have recently emerged with new applications ranging from discovering novel molecular structures to modeling social networks. This paper conducts a comprehensive survey on deep learning-based graph generation approaches and classifies them into five broad categories, namely, autoregressive, autoencoder-based, RL-based, adversarial, and flow-based graph generators, providing the readers a detailed description of the methods in each class. We also present publicly available source codes, commonly used datasets, and the most widely utilized evaluation metrics. Finally, we highlight the existing challenges and discuss future research directions.