Papers

Learning effective representations of entities and relations for knowledge graphs (KGs) is critical to the success of many multi-relational learning tasks. Existing methods based on graph neural networks learn a deterministic embedding function, which lacks sufficient flexibility to explore better choices when dealing with the imperfect and noisy KGs such as the scarce labeled nodes and noisy graph structure. To this end, we propose a novel multi-relational graph Gaussian Process network (GGPN), which aims to improve the flexibility of deterministic methods by simultaneously learning a family of embedding functions, i.e., a stochastic embedding function. Specifically, a Bayesian Gaussian Process (GP) is proposed to model the distribution of this stochastic function and the resulting representations are obtained by aggregating stochastic function values, i.e., messages, from neighboring entities. The two problems incurred when leveraging GP in GGPN are the proper choice of kernel function and the cubic computational complexity. To address the first problem, we further propose a novel kernel function that can explicitly take the diverse relations between each pair of entities into account and be adaptively learned in a data-driven way. We address the second problem by reformulating GP as a Bayesian linear model, resulting in a linear computational complexity. With these two solutions, our GGPN can be efficiently trained in an end-to-end manner. We evaluate our GGPN in link prediction and entity classification tasks, and the experimental results demonstrate the superiority of our method.

poster session 1 @ blue 2, poster session 11 @ blue 2, poster session 1, poster session 11

Across multiple domains from computer vision to speech recognition, machine learning models have been shown to match or outperform human experts at recognition tasks. We lack such a comparison point for Entity Linking. We construct a human benchmark on two standard datasets (TAC KBP 2010 and AIDA (YAGO)) to measure human accuracy. We find that current systems still fall short of human performance. We present DeepType 2, a novel entity linking system that closes the gap. Our proposed approach overcomes shortcomings of previous type-based entity linking systems, and does not use pre-trained language models to reach this level. Three key innovations are responsible for DeepType 2's performance: 1) an abstracted representation of entities that favors shared learning and greater sample efficiency, 2) autoregressive entity features indicating type interactions (e.g. list type homogeneity, shared employers, geographical co-occurrence) with previous predictions that enable globally coherent document-wide predictions, 3) the entire model is trained end to end using a single entity-level maximum likelihood objective function. This is made possible by associating a context-specific score to each of the entity's abstract representation's sub-components (types), and summing these scores to form a candidate entity logit. In this paper, we explain how this factorization focuses the learning on the salient types of the candidate entities. Furthermore, we show how the scores can serve as a rationale for predictions. The key contributions of this work are twofold: 1) we create the first human performance benchmark on standard benchmarks in entity linking (TAC KBP 2010 and AIDA (YAGO)) which will be made publicly available to support further analyses, 2) we obtain a new state of the art on these datasets and are the first to outperform humans on our benchmark. We perform model ablations to measure the contribution of the different facets of our system. We also include an analysis of human and algorithmic errors to provide insights into the causes, notably originating from journalistic style and historical context.

poster session 1 @ blue 2, poster session 11 @ blue 2, oral session 1 @ blue 2, poster session 1, poster session 11, oral session 1

Bugs in source files (SFs) may cause software malfunction, inconveniencing users and even leading to catastrophic accidents. Therefore, the bugs in SFs should be found and fixed quickly. However, from hundreds of candidate SFs, finding buggy SFs is tedious and time consuming. To lessen the burden on developers, deep learning-based bug localization (DLBL) tools can be utilized. Text terms in bug reports and SFs play an important role. However, some terms provide incorrect information and degrade bug localization performance. Therefore, those terms are defined here as "misguiding terms," and an explainable-artificial-intelligence-based identification method is proposed. The effectiveness of the proposed method for DLBL was investigated. When misguiding terms were removed, the mean average precision of the bug localization model improved by 33% on average.

poster session 1 @ blue 2, poster session 11 @ blue 2, poster session 1, poster session 11

Transformer models have demonstrated their promising potential and achieved excellent performance on a series of computer vision tasks. However, the huge computational cost of vision transformers hinders their deployment and application to edge devices. Recent works have proposed to find and remove the unimportant units of vision transformers. Despite achieving remarkable results, these methods take one dimension of network width into consideration and ignore network depth, which is another important dimension for pruning vision transformers. Therefore, we propose a Width & Depth Pruning (WDPruning) framework that reduces both width and depth dimensions simultaneously. Specifically, for width pruning, a set of learnable pruning-related parameters is used to adaptively adjust the width of transformer. For depth pruning, we introduce several shallow classifiers by using the intermediate information of the transformer blocks, which allows images to be classified by shallow classifiers instead of the deeper classifiers. In the inference period, all of the blocks after shallow classifiers can be dropped so they don’t bring additional parameters and computation. Experimental results on benchmark datasets demonstrate that the proposed method can significantly reduce the computational costs of mainstream vision transformers such as DeiT and Swin Transformer with a minor accuracy drop. In particular, on ILSVRC-12, we achieve over 22% pruning ratio of FLOPs by compressing DeiT-Base, even with an increase of 0.14% Top-1 accuracy.

poster session 1 @ blue 2, poster session 11 @ blue 2, poster session 1, poster session 11

Knowledge distillation (KD) has recently emerged as a powerful strategy to transfer knowledge from a pre-trained teacher model to a lightweight student, and has demonstrated its unprecedented success over a wide spectrum of applications. In spite of the encouraging results, the KD process \emph{per se} poses a potential threat to network ownership protection, since the knowledge contained in network can be effortlessly distilled and hence exposed to a malicious user. In this paper, we propose a novel framework, termed as Safe Distillation Box~(SDB), that allows us to wrap a pre-trained model in a virtual box for intellectual property protection. Specifically, SDB preserves the inference capability of the wrapped model to all users, but precludes KD from unauthorized users. For authorized users, on the other hand, SDB carries out a knowledge augmentation scheme to strengthen the KD performances and the results of the student model. In other words, all users may employ a model in SDB for inference, but only authorized users get access to KD from the model. The proposed SDB imposes no constraints over the model architecture, and may readily serve as a plug-and-play solution to protect the ownership of a pre-trained network. Experiments across various datasets and architectures demonstrate that, with SDB, the performance of an unauthorized KD drops significantly while that of an authorized gets enhanced, demonstrating the effectiveness of SDB.

poster session 1 @ blue 2, poster session 11 @ blue 2, poster session 1, poster session 11

Spiking neural networks (SNNs) can be run on neuromorphic devices with ultra-high speed and ultra-low energy consumption because of their binary and event-driven nature. Therefore, SNNs are expected to have various applications, including as generative models being running on edge devices to create high-quality images. In this study, we build a variational autoencoder (VAE) with SNN to enable image generation. VAE is known for its stability among generative models; recently, its quality advanced. In vanilla VAE, the latent space is represented as a normal distribution, and floating-point calculations are required in sampling. However, this is not possible in SNNs because all features must be binary time series data. Therefore, we constructed the latent space with an autoregressive SNN model, and randomly selected samples from its output to sample the latent variables. This allows the latent variables to follow the Bernoulli process and allows variational learning. Thus, we build the Fully Spiking Variational Autoencoder where all modules are constructed with SNN. To the best of our knowledge, we are the first to build a VAE only with SNN layers. We experimented with several datasets, and confirmed that it can generate images with the same or better quality compared to conventional ANNs.

poster session 1 @ blue 2, poster session 11 @ blue 2, poster session 1, poster session 11

This work introduces Fractional Activation Linear Units (FALUs), a flexible generalization of adaptive activation functions. Leveraging principles from fractional calculus, FALUs define a diverse family of activation functions that encompass many traditional and state-of-the-art activation functions, including the Sigmoid, ReLU, Swish, GELU, and Gaussian functions, as well as a large variety of smooth interpolations between these functions. Our technique introduces only a small number of additional trainable parameters, and requires no additional specialized optimization or initialization procedures. For this reason, FALUs present a seamless and rich automated solution to the problem of activation function optimization. Through experiments on a variety of conventional tasks and network architectures, we demonstrate the effectiveness of FALUs when compared to traditional and state-of-the-art activation functions.

poster session 1 @ blue 2, poster session 11 @ blue 2, poster session 1, poster session 11

The Mixup scheme suggests mixing a pair of samples to create an augmented training sample and has gained considerable attention recently for improving the generalizability of neural networks. A straightforward and widely used extension of Mixup is to combine with regional dropout-like methods: removing random patches from a sample and replacing it with the features from another sample. Albeit their simplicity and effectiveness, these methods are prone to create harmful samples due to their randomness. To address this issue, 'maximum saliency' strategies were recently proposed: they select only the most informative features to prevent such a phenomenon. However, they now suffer from lack of sample diversification as they always deterministically select regions with maximum saliency, injecting bias into the augmented data. In this paper, we present, a novel, yet simple Mixup-variant that captures the best of both worlds. Our idea is two-fold. By stochastically sampling the features and ‘grafting’ them onto another sample, our method effectively generates diverse yet meaningful samples. Its second ingredient is to produce the label of the grafted sample by mixing the labels in a saliency-calibrated fashion, which rectifies supervision misguidance introduced by the random sampling procedure. Our experiments under CIFAR, Tiny-ImageNet, and ImageNet datasets show that our scheme outperforms the current state-of-the-art augmentation strategies not only in terms of classification accuracy, but is also superior in coping under stress conditions such as data corruption and object occlusion.

poster session 1 @ blue 2, poster session 11 @ blue 2, oral session 11 @ blue 2, poster session 1, poster session 11, oral session 11

The conventional wisdom behind learning deep classification models is to focus on bad-classified examples and ignore well-classified examples that are far from the decision boundary. For instance, when training with cross-entropy loss, examples with higher likelihoods (i.e., well-classified examples) contribute smaller gradients in back-propagation. However, we theoretically show that this common practice hinders representation learning, energy optimization, and margin growth. To counteract this deficiency, we propose to reward well-classified examples with additive bonuses to revive their contribution to the learning process. This counterexample theoretically addresses these three issues. We empirically support this claim by directly verifying the theoretical results or significant performance improvement with our counterexample on diverse tasks, including image classification, graph classification, and machine translation. Furthermore, this paper shows that we can deal with complex scenarios, such as imbalanced classification, OOD detection, and applications under adversarial attacks, because our idea can solve these three issues.Code is available at \url{https://github.com/lancopku/well-classified-examples-are-underestimated}.

poster session 1 @ blue 2, poster session 11 @ blue 2, poster session 1, poster session 11

In this paper, we propose Deep amortized Relational Model (DaRM) with group-wise hierarchical generative process for community discovery and link prediction on relational data (e.g., graph, network). It provides an efficient neural relational model architecture by grouping nodes in a group-wise view rather than node-wise or edge-wise view. DaRM simultaneously learns what makes a group, how to divide nodes into groups, and how to adaptively control the number of groups. The dedicated group generative process is able to sufficiently exploit pair-wise or higher-order interactions between data points in both inter-group and intra-group, which is useful to sufficiently mine the hidden structure among data. A series of experiments have been conducted on both synthetic and real-world datasets. The experimental results demonstrated that DaRM can obtain high performance on both community detection and link prediction tasks.

poster session 1 @ blue 2, poster session 11 @ blue 2, poster session 1, poster session 11

Lack of explainability in artificial intelligence, specifically deep neural networks, remains a bottleneck for implementing models in practice. Popular techniques such as Gradient-weighted Class Activation Mapping (Grad-CAM) provide a coarse map of salient features in an image, which rarely tells the whole story of what a convolutional neural network(CNN) learned. Using COVID-19 chest X-rays, we present a method for interpreting what a CNN has learned by utilizing Generative Adversarial Networks (GANs). Our GAN framework disentangles lung structure from COVID-19 features. Using this GAN, we can visualize the transition of a pair of COVID negative lungs in a chest radiograph to a COVID positive pair by interpolating in the latent space of the GAN, which provides fine-grained visualization of how the CNN responds to varying features within the lungs.

poster session 1 @ blue 2, poster session 11 @ blue 2, poster session 1, poster session 11

Transformers achieve state-of-the-art performance for natural language processing tasks by pre-training on large-scale text corpora. They are extremely compute-intensive and have very high sample complexity. Memory replay is a mechanism that remembers and reuses past examples by saving to and replaying from a memory buffer. It has been successfully used in reinforcement learning and GANs due to better sample efficiency. In this paper, we propose Transformer with Memory Replay, which integrates memory replay with transformer, making transformer more sample efficient. Experiments on GLUE and SQuAD benchmark datasets showed that Transformer with Memory Replay can achieve at least 1% point increase compared to the baseline transformer model when pre-trained with the same number of examples. Further, by adopting a careful design that reduces the wall-clock time overhead of memory replay, we also empirically achieve a better runtime efficiency.

poster session 1 @ blue 2, poster session 11 @ blue 2, poster session 1, poster session 11

Long Short-Term Memory (LSTM) and Transformers are two popular neural architectures used for natural language processing tasks. Theoretical results show that both are Turing-complete and can represent any context-free language (CFL).In practice, it is often observed that Transformer models have better representation power than LSTM. But the reason is barely understood. We study such practical differences between LSTM and Transformer and propose an explanation based on their latent space decomposition patterns. To achieve this goal, we introduce an oracle training paradigm, which forces the decomposition of the latent representation of LSTMand the Transformer and supervises with the transitions of the Pushdown Automaton (PDA) of the corresponding CFL. With the forced decomposition, we show that the performance upper bounds of LSTM and Transformer in learning CFL are close: both of them can simulate a stack and perform stack operation along with state transitions. However, the absence of forced decomposition leads to the failure of LSTM models to capture the stack and stack operations, while having a marginal impact on the Transformer model. Lastly, we connect the experiment on the prototypical PDA to a real-world parsing task to re-verify the conclusions

poster session 1 @ blue 2, poster session 11 @ blue 2, oral session 1 @ blue 2, poster session 1, poster session 11, oral session 1

A promising trend in deep learning replaces traditional feedforward networks with implicit networks. Unlike traditional networks, implicit networks solve a fixed point equation to compute inferences. Solving for the fixed point varies in complexity, depending on provided data and an error tolerance. Importantly, implicit networks may be trained with fixed memory costs in stark contrast to feedforward networks, whose memory requirements scale linearly with depth. However, there is no free lunch --- backpropagation through implicit networks often requires solving a costly Jacobian-based equation arising from the implicit function theorem. We propose Jacobian-Free Backpropagation (JFB), a fixed-memory approach that circumvents the need to solve Jacobian-based equations. JFB makes implicit networks faster to train and significantly easier to implement, without sacrificing test accuracy. Our experiments show implicit networks trained with JFB are competitive with feedforward networks and prior implicit networks given the same number of parameters.

poster session 1 @ blue 2, poster session 11 @ blue 2, poster session 1, poster session 11

Spatiotemporal graph represents a crucial data structure where the nodes and edges are embedded in a geometric space and their attribute values can evolve dynamically over time. Nowadays, spatiotemporal graph data is becoming increasingly popular and important, ranging from microscale (e.g. protein folding), to middle-scale (e.g. dynamic functional connectivity), to macro-scale (e.g. human mobility network). Although disentangling and understanding the correlations among spatial, temporal, and graph aspects have been a long-standing key topic in network science, they typically rely on network processes hypothesized by human knowledge. They usually fit well towards the properties that the predefined principles are tailored for, but usually cannot do well for the others, especially for many key domains where the human has yet very limited knowledge such as protein folding and biological neuronal networks. In this paper, we aim at pushing forward the modeling and understanding of spatiotemporal graphs via new disentangled deep generative models. Specifically, a new Bayesian model is proposed that factorizes spatiotemporal graphs into spatial, temporal, and graph factors as well as the factors that explain the interplay among them. A variational objective function and new mutual information thresholding algorithms driven by information bottleneck theory have been proposed to maximize the disentanglement among the factors with theoretical guarantees. Qualitative and quantitative experiments on both synthetic and real-world datasets demonstrate the superiority of the proposed model over the state-of-the-arts by up to 69.2\% for graph generation and 41.5\% for interpretability.

poster session 1 @ blue 2, poster session 11 @ blue 2, oral session 11 @ blue 2, poster session 1, poster session 11, oral session 11

Normalizing flows have shown to be a promising approach to deep generative modeling due to their ability to exactly evaluate density --- other alternatives either implicitly model the density or use approximate surrogate density. In this work, we present a differentially private normalizing flow model for heterogeneous tabular data. Normalizing flows are in general not amenable to differentially private training because they require complex neural networks with larger depth (compared to other generative models) and use specialized architectures for which per-example gradient computation is difficult (or unknown). To reduce the parameter complexity, the proposed model introduces a conditional spline flow which simulates transformations at different stages depending on additional input and is shared among sub-flows. For privacy, we introduce two fine-grained gradient clipping strategies that provide a better signal-to-noise ratio and derive fast gradient clipping methods for layers with custom parameterization. Our empirical evaluations show that the proposed model preserves statistical properties of original dataset better than other baselines.

poster session 1 @ blue 2, poster session 11 @ blue 2, poster session 1, poster session 11

Sparse Neural Networks (NNs) can match the generalization of dense NNs using a fraction of the compute/storage for inference, and have the potential to enable efficient training. However, naively training unstructured sparse NNs from random initialization results in significantly worse generalization, with the notable exceptions of Lottery Tickets (LTs) and Dynamic Sparse Training (DST). In this work, we attempt to answer: (1) why training unstructured sparse networks from random initialization performs poorly and; (2) what makes LTs and DST the exceptions? We show that sparse NNs have poor gradient flow at initialization and propose a modified initialization for unstructured connectivity. Furthermore, we find that DST methods significantly improve gradient flow during training over traditional sparse training methods. Finally, we show that LTs do not improve gradient flow, rather their success lies in re-learning the pruning solution they are derived from — however, this comes at the cost of learning novel solutions.

poster session 1 @ blue 2, poster session 11 @ blue 2, oral session 1 @ blue 2, poster session 1, poster session 11, oral session 1

We introduce Monte-Carlo Attention (MCA), a randomized approximation method for reducing the computational cost of self-attention mechanisms in Transformer architectures. MCA exploits the fact that the importance of each token in an input sequence vary with respect to their attention scores; thus, some degree of error can be tolerable when encoding tokens with low attention. Using approximate matrix multiplication, MCA applies different error bounds to encode input tokens such that those with low attention scores are computed with relaxed precision, whereas errors of salient elements are minimized. MCA can operate in parallel with other attention optimization schemes and does not require model modification. We study the theoretical error bounds and demonstrate that MCA reduces attention complexity (in FLOPS) for various Transformer models by up to 11x in GLUE benchmarks without compromising model accuracy.

poster session 1 @ blue 2, poster session 11 @ blue 2, poster session 1, poster session 11

Variational Autoencoders (VAEs) suffer from degenerated performance, when learning several successive tasks. This is caused by catastrophic forgetting. In order to address the knowledge loss, VAEs are using either Generative Replay (GR) mechanisms or Expanding Network Architectures (ENA). In this paper we study the forgetting behaviour of VAEs using a joint GR and ENA methodology, by deriving an upper bound on the negative marginal log-likelihood. This theoretical analysis provides new insights into how VAEs forget the previously learnt knowledge during lifelong learning. The analysis indicates the best performance achieved when considering model mixtures, under the ENA framework, where there are no restrictions on the number of components. However, an ENA-based approach may require an excessive number of parameters. This motivates us to propose a novel Dynamic Expansion Graph Model (DEGM). DEGM expands its architecture, according to the novelty associated with each new databases, when compared to the information already learnt by the network from previous tasks. DEGM training optimizes knowledge structuring, characterizing the joint probabilistic representations corresponding to the past and more recently learned tasks. We demonstrate that DEGM guarantees optimal performance for each task while also minimizing the required number of parameters. Supplementary materials (SM) and source code are available in https://github.com/dtuzi123/Expansion-Graph-Model.

poster session 1 @ blue 2, poster session 11 @ blue 2, oral session 11 @ blue 2, poster session 1, poster session 11, oral session 11

We introduce three new generative models for time series that are based on Euler discretization of Stochastic Differential Equations (SDEs) and Wasserstein metrics. Two of these methods rely on the adaptation of generative adversarial networks (GANs) to time series. The third algorithm, called Conditional Euler Generator (CEGEN), minimizes a dedicated distance between the transition probability distributions over all time steps. In the context of Itô processes, we provide theoretical guarantees that minimizing this criterion implies accurate estimations of the drift and volatility parameters. Empirically, CEGEN outperforms state-of-the-art and GANs on both marginal and temporal dynamic metrics. Besides, correlation structures are accurately identified in high dimension. When few real data points are available, we verify the effectiveness of CEGEN when combined with transfer learning methods on model-based simulations. Finally, we illustrate the robustness of our methods on various real-world data sets.

poster session 1 @ blue 2, poster session 11 @ blue 2, poster session 1, poster session 11

Understanding the capability of Generative Adversarial Networks (GANs) in learning the full spectrum of spatial frequencies, that is, beyond the low-frequency dominant spectrum of natural images, is critical for assessing the reliability of GAN-generated data in any detail-sensitive application. In this work, we show that the ability of convolutional GANs to learn an image distribution depends on the spatial frequency of the underlying carrier signal, that is, they have a bias against learning high spatial frequencies. Our findings are consistent with the recent observations of high-frequency artifacts in GAN-generated images, but further suggest that such artifacts are the consequence of an underlying bias. We also provide a theoretical explanation for this bias as the manifestation of linear dependencies present in the spectrum of filters of a typical generative Convolutional Neural Network (CNN). Finally, by proposing a proof-of-concept method that can effectively manipulate this bias towards other spatial frequencies, we show that the bias is not fixed and can be exploited to explicitly direct computational resources towards any specific spatial frequency of interest in a dataset, with minimal computational overhead.

poster session 1 @ blue 2, poster session 11 @ blue 2, poster session 1, poster session 11

In contexts where debate and deliberation is the norm, participants are regularly presented with new information that conflicts with their original beliefs. When required to update their beliefs (belief alignment), they may choose arguments that align with their worldview (confirmation bias). We test this and competing hypotheses in a constraint-based modeling approach to predict the winning arguments in multi-party interactions in the Reddit ChangeMyView dataset. We impose structural constraints that reflect competing hypotheses on a hierarchical generative Variational Auto-encoder. Our findings suggest that when arguments are further from the initial belief state of the target, they are more likely to succeed.

poster session 1 @ blue 2, poster session 11 @ blue 2, poster session 1, poster session 11

This abstract proposes an approach towards goal-oriented modeling of the detection and modeling complex social phenomena in multiparty discourse in an online political strategy game. We developed a two-tier approach that first encodes sociolinguistic behavior as linguistic features then use reinforcement learning to estimate the advantage afforded to any player. In the first tier, sociolinguistic behavior, such as Friendship and Reasoning, that speakers use to influence others are encoded as linguistic features to identify the persuasive strategies applied by each player in simultaneous two-party dialogues. In the second tier, a reinforcement learning approach is used to estimate a graph-aware reward function to quantify the advantage afforded to each player based on their standing in this multiparty setup. We apply this technique to the game Diplomacy, using a dataset comprising of over 15,000 messages exchanged between 78 users. Our graph-aware approach shows robust performance compared to a context-agnostic setup.

poster session 1 @ blue 2, poster session 11 @ blue 2, poster session 1, poster session 11

Monaural speech enhancement (SE) at an extremely low signal-to-noise ratio (SNR) condition is a challenging problem and rarely investigated in previous studies. Most SE methods experience failures in this situation due to three major factors: overwhelmed vocals, expanded SNR range, and short-sighted feature processing modules. In this paper, we present a novel and general training paradigm dubbed repetitive learning (RL). Unlike curriculum learning that focuses on learning multiple different tasks sequentially, RL is more inclined to learn the same content repeatedly where the knowledge acquired in previous stages can be used to facilitate calibrating feature representations. We further propose an RL-based end-to-end SE method named SERL. Experimental results on TIMIT dataset validate the superior performance of our method.

poster session 1 @ blue 2, poster session 11 @ blue 2, poster session 1, poster session 11