AAAI 2024 Papers

Self-supervised approaches for video have shown impressive results in video understanding tasks. However, unlike early works that leverage temporal self-supervision, current state-of-the-art methods primarily rely on tasks from the image domain (e.g., contrastive learning) that do not explicitly promote the learning of temporal features. We identify two factors that limit existing temporal self-supervision: 1) tasks are too simple, resulting in saturated training performance, and 2) we uncover shortcuts based on local appearance statistics that hinder the learning of high-level features. To address these issues, we propose 1) a more challenging reformulation of temporal self-supervision as frame-level (rather than clip-level) recognition tasks and 2) an effective augmentation strategy to mitigate shortcuts. Our model extends a representation of single video frames, pre-trained through contrastive learning, with a transformer that we train through temporal self-supervision. We demonstrate experimentally that our more challenging frame-level task formulations and the removal of shortcuts drastically improve the quality of features learned through temporal self-supervision. Our extensive experiments show state-of-the-art performance across 10 video understanding datasets, illustrating the generalization ability and robustness of our learned video representations. Project Page: https://daveishan.github.io/nms-webpage.

Internal regret is a measure of performance of an online learning algorithm, which measures the change in performance by substituting every occurrence of a given action i by an alternative action j. Algorithms for minimizing internal regret are known for the finite experts setting, including a general reduction to the problem of minimizing external regret for this case. The reduction however crucially depends on the finiteness of the action space. In this work we approach the problem of minimizing internal regret for a continuous action space. For the full information setting, we show how to obtain O(sqrt(T)) internal regret for the class of Lipschitz functions, as well as non-Lipschitz dispersed functions, i.e. the non-Lipschitzness may not concentrate in a small region of the action space. We also consider extensions to partial feedback settings, and again obtain sublinear internal regret. Finally we discuss applications of internal regret minimization over continuous spaces to correlated equilibria in pricing problems and auction design, as well as to data-driven hyperparameter tuning.

Knowledge distillation aims at reducing model size without compromising much performance. Recent work has applied it to large vision-language (VL) Transformers, and has shown that attention maps in the multi-head attention modules of vision-language Transformers contain extensive intra-modal and cross-modal co-reference relations to be distilled. The standard approach is to apply a one-to-one attention map distillation loss, i.e. the Teacher's first attention head instructs the Student's first head, the second teaches the second, and so forth, but this only works when the numbers of attention heads in the Teacher and Student are the same. To remove this constraint, we propose a new Attention Map Alignment Distillation (AMAD) method for Transformers with multi-head attention, which works for a Teacher and a Student with different numbers of attention heads. Specifically, we soft-align different heads in Teacher and Student attention maps using a cosine similarity weighting. The Teacher head contributes more to the Student heads for which it has a higher similarity weight. Each Teacher head contributes to all the Student heads by minimizing the divergence between the attention activation distributions for the soft-aligned heads. No head is left behind. This distillation approach operates like cross-attention. We experiment on distilling VL-T5 and BLIP, and apply AMAD loss on their T5, BERT, and ViT sub-modules. We show, under vision-language setting, that AMAD outperforms conventional distillation methods on VQA-2.0, COCO captioning, and Multi30K translation datasets. We further show that even without VL pre-training, the distilled VL-T5 models outperform corresponding VL pre-trained VL-T5 models that are further fine-tuned by ground-truth signals, and that fine-tuning distillation can also compensate to some degree for the absence of VL pre-training for BLIP models.

Existing deep-learning-based methods for nighttime video deraining rely on synthetic data due to the absence of real-world paired data. However, the intricacies of the real world, particularly with the presence of light effects and low-light regions affected by noise, create significant domain gaps, hampering synthetic-trained models in removing rain streaks properly and leading to over-saturation and color shifts. Motivated by this, we introduce NightRain, a novel nighttime video deraining method with adaptive-rain-removal and adaptive-correction. Our adaptive-rain-removal uses unlabeled rain videos to enable our model to derain real-world rain videos, particularly in regions affected by complex light effects. The idea is to allow our model to obtain rain-free regions based on the confidence scores. Once rain-free regions and the corresponding regions from our input are obtained, we can have region-based paired real data. These paired data are used to train our model using a teacher-student framework, allowing the model to iteratively learn from less challenging regions to more challenging regions. Our adaptive-correction aims to rectify errors in our model's predictions, such as over-saturation and color shifts. The idea is to learn from clear night input training videos based on the differences or distance between those input videos and their corresponding predictions. Our model learns from these differences, compelling our model to correct the errors. From extensive experiments, our method demonstrates state-of-the-art performance. It achieves a PSNR of 26.73dB, surpassing existing nighttime video deraining methods by a substantial margin of 13.7%.

We consider the sparse contextual bandit problem where arm feature affects reward through the inner product of sparse parameters. Recent studies have developed sparsity-agnostic algorithms based on the greedy arm selection policy. However, the analysis of these algorithms requires strong assumptions on the arm feature distribution to ensure that the greedily selected samples are sufficiently diverse; One of the most common assumptions, relaxed symmetry, imposes approximate origin-symmetry on the distribution, which cannot allow distributions that has origin-asymmetric support. In this paper, we show that the greedy algorithm is applicable to a wider range of the arm feature distributions from two aspects. Firstly, we show that a mixture distribution that has a greedy-applicable component is also greedy-applicable. Second, we propose new distribution classes, related to Gaussian mixture, discrete, and radial distribution, for which the sample diversity is guaranteed. The proposed classes can describe distributions with origin-asymmetric support and, in conjunction with the first claim, provide theoretical guarantees of the greedy policy for a very wide range of the arm feature distributions.

Convolutional neural networks-based video de-raining methods commonly rely on dense intensity frames captured by CMOS sensors. However, the limited temporal resolution of these sensors hinders the capture of dynamic rainfall information, limiting further improvement in de-raining performance. This study aims to overcome this issue by incorporating the neuromorphic event signal into the video de-raining to enhance the dynamic information perception. Specifically, we first utilize the dynamic information from the event signal as prior knowledge, and integrate it into existing de-raining objectives to better constrain the solution space. We then design an optimization algorithm to solve the objective, and construct a de-raining network with CNNs as the backbone architecture using a modular strategy to mimic the optimization process. To further explore the temporal correlation of the event signal, we incorporate a spiking self-attention module into our network. By leveraging the low latency and high temporal resolution of the event signal, along with the spatial and temporal representation capabilities of convolutional and spiking neural networks, our model captures more accurate dynamic information and significantly improves de-raining performance. For example, our network achieves a 1.24dB improvement on the SynHeavy25 dataset compared to the previous state-of-the-art method, while utilizing only 39% of the parameters.

Evolutionary Algorithms (EA) have been leveraged to tackle the challenges faced while using GANs such as mode collapse, vanishing gradient, latent space search, etc. However, the existing techniques of using EA with GANs operate backpropagation and EA in isolation from each other, leaving ample room for further exploration. This paper creates a collaborative bridge between EA and GANs by exploring a neuroevolution method for utilising both EA and backpropagation-based optimisation, simultaneously, for a multi-generator GAN architecture. Experiments conducted using a standard dataset with variants of the proposed method highlight the towering impact of each of the components involved in the proposed method.

The goal of this thesis is to address knowledge graph completion tasks using neuro-symbolic methods. Neuro-symbolic methods allow the joint utilization of symbolic information defined as meta-rules in ontologies and knowledge graph embedding methods that represent entities and relations of the graph in a low-dimensional vector space. This approach has the potential to improve the resolution of knowledge graph completion tasks in terms of reliability, interpretability, data-efficiency and robustness.

Optimal control deals with optimization problems in which variables steer a dynamical system, and its outcome contributes to the objective function. Two classical approaches to solving these problems are Dynamic Programming and the Pontryagin Maximum Principle. In both approaches, Hamiltonian equations offer an interpretation of optimality through auxiliary variables known as costates. However, Hamiltonian equations are rarely used due to their reliance on forward-backward algorithms across the entire temporal domain. This paper introduces a novel neural-based approach to optimal control. Neural networks are employed not only for implementing state dynamics but also for estimating costate variables. The parameters of the latter network are determined at each time step using a newly introduced local policy referred to as the time-reversed generalized Riccati equation. This policy is inspired by a result discussed in the Linear Quadratic (LQ) problem, which we conjecture stabilizes state dynamics. We support this conjecture by discussing experimental results from a range of optimal control case studies.

We propose the Intuitive Reasoning Network (IRENE) - a novel neural model for intuitive psychological reasoning about agents' goals, preferences, and actions that can generalise previous experiences to new situations. IRENE combines a graph neural network for learning agent and world state representations with a transformer to encode the task context. When evaluated on the challenging Baby Intuitions Benchmark, IRENE achieves new state-of-the-art performance on three out of its five tasks - with up to 48.9% improvement. In contrast to existing methods, IRENE is able to bind preferences to specific agents, to better distinguish between rational and irrational agents, and to better understand the role of blocking obstacles. We also investigate, for the first time, the influence of the training tasks on test performance. Our analyses demonstrate the effectiveness of IRENE in combining prior knowledge gained during training for unseen evaluation tasks.

We explore the generalization of the implicit representation in the physical simulation task. Traditional time-dependent partial differential equations (PDEs) solvers for physical simulation often adopt the grid or mesh for spatial discretization, which is memory-consuming for high resolution and lack of adaptivity. Many implicit representations like local extreme machine or Siren are proposed but they are still too compact to suffer from limited accuracy in handling local details and a long time of convergence. We contribute a neural simulation framework based on multi-resolution hash grid representation to introduce hierarchical consideration of global and local information, simultaneously. Furthermore, we propose two key strategies: 1) a numerical gradient method for computing high-order derivatives with boundary conditions; 2) a range analysis sample method for fast neural geometry boundary sampling with dynamic topologies. Our method shows much higher accuracy and strong flexibility for various simulation problems: e.g., large elastic deformations, complex fluid dynamics, and multi-scale phenomena which remain challenging for existing neural physical solvers.

A primary challenge of physics-informed machine learning (PIML) is its generalization beyond the training domain, especially when dealing with complex physical problems represented by partial differential equations (PDEs). This paper aims to enhance the generalization capabilities of PIML, facilitating practical, real-world applications where accurate predictions in unexplored regions are crucial. We leverage the inherent causality and temporal sequential characteristics of PDE solutions to fuse PIML models with recurrent neural architectures based on systems of ordinary differential equations, referred to as neural oscillators. Through effectively capturing long-time dependencies and mitigating the exploding and vanishing gradient problem, neural oscillators foster improved generalization in PIML tasks. Extensive experimentation involving time-dependent nonlinear PDEs and biharmonic beam equations demonstrates the efficacy of the proposed approach. Incorporating neural oscillators outperforms existing state-of-the-art methods on benchmark problems across various metrics. Consequently, the proposed method improves the generalization capabilities of PIML, providing accurate solutions for extrapolation and prediction beyond the training data.

Probabilistic circuits (PCs) such as sum-product networks efficiently represent large multi-variate probability distributions. They are preferred in practice over other probabilistic representations, such as Bayesian and Markov networks, because PCs can solve marginal inference (MAR) tasks in time that scales linearly in the size of the network. Unfortunately, the most probable explanation (MPE) task and its generalization, the marginal maximum-a-posteriori (MMAP) inference task remain NP-hard in these models. Inspired by the recent work on using neural networks for generating near-optimal solutions to optimization problems such as integer linear programming, we propose an approach that uses neural networks to approximate MMAP inference in PCs. The key idea in our approach is to approximate the cost of an assignment to the query variables using a continuous multilinear function and then use the latter as a loss function. The two main benefits of our new method are that it is self-supervised, and after the neural network is learned, it requires only linear time to output a solution. We evaluate our new approach on several benchmark datasets and show that it outperforms three competing linear time approximations: max-product inference, max-marginal inference, and sequential estimation, which are used in practice to solve MMAP tasks in PCs.

Deep reinforcement learning (RL) has shown remarkable success in specific offline decision-making scenarios, yet its theoretical guarantees are still under development. Existing works on offline RL theory primarily emphasize a few trivial settings, such as linear MDP or general function approximation with strong assumptions and independent data, which lack guidance for practical use. The coupling of deep learning and Bellman residuals makes this problem challenging, in addition to the difficulty of data dependence. In this paper, we establish a non-asymptotic estimation error of pessimistic offline RL using general neural network approximation with C-mixing data regarding the structure of networks, the dimension of datasets, and the concentrability of data coverage, under mild assumptions. Our result shows that the estimation error consists of two parts: the first converges to zero at a desired rate on the sample size with partially controllable concentrability, and the second becomes negligible if the residual constraint is tight. This result demonstrates the explicit efficiency of deep adversarial offline RL frameworks. We utilize the empirical process tool for C-mixing sequences and the neural network approximation theory for the Holder class to achieve this. We also develop methods to bound the Bellman estimation error caused by function approximation with empirical Bellman constraint perturbations. Additionally, we present a result that lessens the curse of dimensionality using data with low intrinsic dimensionality and function classes with low complexity. Our estimation provides valuable insights into the development of deep offline RL and guidance for algorithm model design.

Graph Structure Learning (GSL) has demonstrated considerable potential in the analysis of graph-unknown non-Euclidean data across a wide range of domains. However, constructing an end-to-end graph structure learning model poses a challenge due to the impediment of gradient flow caused by the nearest neighbor sampling strategy. In this paper, we construct a differential graph structure learning model by replacing the non-differentiable nearest neighbor sampling with a differentiable sampling using the reparameterization trick. Under this framework, we argue that the act of sampling nearest neighbors may not invariably be essential, particularly in instances where node features exhibit a significant degree of similarity. To alleviate this issue, the bell-shaped Gaussian Similarity (GauSim) modeling is proposed to sample non-nearest neighbors. To adaptively model the similarity, we further propose Neural Gaussian Similarity (NeuralGauSim) with learnable parameters featuring flexible sampling behaviors. In addition, we develop a scalable method by transferring the large-scale graph to the transition graph to significantly reduce the complexity. Experimental results demonstrate the effectiveness of the proposed methods.

Biological sequence nearest neighbor search plays a fundamental role in bioinformatics. To alleviate the pain of quadratic complexity for conventional distance computation, neural distance embeddings, which project sequences into geometric space, have been recognized as a promising paradigm. To maintain the distance order between sequences, these models all deploy triplet loss and use intuitive methods to select a subset of triplets for training from a vast selection space. However, we observed that such training often enables models to distinguish only a fraction of distance orders, leaving others unrecognized. Moreover, naively selecting more triplets for training under the state-of-the-art network not only adds costs but also hampers model performance. In this paper, we introduce Bio-kNN: a kNN search framework for biological sequences. It includes a systematic triplet selection method and a multi-head network, enhancing the discernment of all distance orders without increasing training expenses. Initially, we propose a clustering-based approach to partition all triplets into several clusters with similar properties, and then select triplets from these clusters using an innovative strategy. Meanwhile, we noticed that simultaneously training different types of triplets in the same network cannot achieve the expected performance, thus we propose a multi-head network to tackle this. Our network employs a convolutional neural network(CNN) to extract local features shared by all clusters, and then learns a multi-layer perception(MLP) head for each cluster separately. Besides, we treat CNN as a special head, thereby integrating crucial local features which are neglected in previous models into our model for similarity recognition. Extensive experiments show that our Bio-kNN significantly outperforms the state-of-the-art methods on two large-scale datasets without increasing the training cost.

Notions of transition invariants and closure certificates have seen recent use in the formal verification of controlled dynamical systems against \omega-regular properties. Unfortunately, existing approaches face limitations in two directions. First, they require a closed-form mathematical expression representing the model of the system. Such an expression may be difficult to find, too complex to be of any use, or unavailable due to security or privacy constraints. Second, finding such invariants typically rely on optimization techniques such as sum-of-squares (SOS) or satisfiability modulo theory (SMT) solvers. This restricts the classes of systems that need to be formally verified. To address these drawbacks, we introduce a notion of neural closure certificates. We present a data-driven algorithm that trains a neural network to represent a closure certificate. Our approach is formally correct under some mild assumptions, i.e., one is able to formally show that the unknown system satisfies the \omega-regular property of interest if a neural closure certificate can be computed. Finally, we demonstrate the efficacy of our approach with relevant case studies.

The ability of humans to understand the world in terms of cause and effect relationships, as well as their ability to compress information into abstract concepts, are two hallmark features of human intelligence. These two topics have been studied in tandem under the theory of causal abstractions, but it is an open problem how to best leverage abstraction theory in real-world causal inference tasks, where the true model is not known, and limited data is available in most practical settings. In this paper, we focus on a family of causal abstractions constructed by clustering variables and their domains, redefining abstractions to be amenable to individual causal distributions. We show that such abstractions can be learned in practice using Neural Causal Models, allowing us to utilize the deep learning toolkit to solve causal tasks (identification, estimation, sampling) at different levels of abstraction granularity. Finally, we show how representation learning can be used to learn abstractions, which we apply in our experiments to scale causal inferences to high dimensional settings such as with image data.

In neural memory decoding, a concept being mentally recalled is identified using brain data. Recently, the feasibility of neural memory decoding with EEG data has been demonstrated. Here we propose a new application – neural information retrieval – that uses neural memory decoding to allow a document to be retrieved merely by thinking about it. In this paper we describe neural memory decoding, define the application of neural information retrieval, present experimental results related to the practicality of the application, and discuss issues of deployment and data privacy.