When mining large datasets in order to predict new data, limitations of the principles behind statistical machine learning pose a serious challenge not only to the Big Data deluge, but also to the traditional assumptions that data generating processes are biased toward low algorithmic complexity. Even when one assumes an underlying algorithmic-informational bias toward simplicity in finite dataset generators, we show that fully automated, with or without access to pseudo-random generators, computable learning algorithms, in particular those of statistical nature used in current approaches to machine learning (including deep learning), can always be deceived, naturally or artificially, by sufficiently large datasets. In particular, we demonstrate that, for every finite learning algorithm, there is a sufficiently large dataset size above which the algorithmic probability of an unpredictable deceiver is an upper bound (up to a multiplicative constant that only depends on the learning algorithm) for the algorithmic probability of any other larger dataset. In other words, very large and complex datasets are as likely to deceive learning algorithms into a "simplicity bubble" as any other particular dataset. These deceiving datasets guarantee that any prediction will diverge from the high-algorithmic-complexity globally optimal solution while converging toward the low-algorithmic-complexity locally optimal solution. We discuss the framework and empirical conditions for circumventing this deceptive phenomenon, moving away from statistical machine learning towards a stronger type of machine learning based on, or motivated by, the intrinsic power of algorithmic information theory and computability theory.
From the original theoretically well-defined spectral graph convolution to the subsequent spatial bassed message-passing model, spatial locality (in vertex domain) acts as a fundamental principle of most graph neural networks (GNNs). In the spectral graph convolution, the filter is approximated by polynomials, where a $k$-order polynomial covers $k$-hop neighbors. In the message-passing, various definitions of neighbors used in aggregations are actually an extensive exploration of the spatial locality information. For learning node representations, the topological distance seems necessary since it characterizes the basic relations between nodes. However, for learning representations of the entire graphs, is it still necessary to hold? In this work, we show that such a principle is not necessary, it hinders most existing GNNs from efficiently encoding graph structures. By removing it, as well as the limitation of polynomial filters, the resulting new architecture significantly boosts performance on learning graph representations. We also study the effects of graph spectrum on signals and interpret various existing improvements as different spectrum smoothing techniques. It serves as a spatial understanding that quantitatively measures the effects of the spectrum to input signals in comparison to the well-known spectral understanding as high/low-pass filters. More importantly, it sheds the light on developing powerful graph representation models.
Despite its great success, deep learning severely suffers from robustness; that is, deep neural networks are very vulnerable to adversarial attacks, even the simplest ones. Inspired by recent advances in brain science, we propose the Denoised Internal Models (DIM), a novel generative autoencoder-based model to tackle this challenge. Simulating the pipeline in the human brain for visual signal processing, DIM adopts a two-stage approach. In the first stage, DIM uses a denoiser to reduce the noise and the dimensions of inputs, reflecting the information pre-processing in the thalamus. Inspired from the sparse coding of memory-related traces in the primary visual cortex, the second stage produces a set of internal models, one for each category. We evaluate DIM over 42 adversarial attacks, showing that DIM effectively defenses against all the attacks and outperforms the SOTA on the overall robustness.
In this paper, we propose a new joint object detection and tracking (JoDT) framework for 3D object detection and tracking based on camera and LiDAR sensors. The proposed method, referred to as 3D DetecTrack, enables the detector and tracker to cooperate to generate a spatio-temporal representation of the camera and LiDAR data, with which 3D object detection and tracking are then performed. The detector constructs the spatio-temporal features via the weighted temporal aggregation of the spatial features obtained by the camera and LiDAR fusion. Then, the detector reconfigures the initial detection results using information from the tracklets maintained up to the previous time step. Based on the spatio-temporal features generated by the detector, the tracker associates the detected objects with previously tracked objects using a graph neural network (GNN). We devise a fully-connected GNN facilitated by a combination of rule-based edge pruning and attention-based edge gating, which exploits both spatial and temporal object contexts to improve tracking performance. The experiments conducted on both KITTI and nuScenes benchmarks demonstrate that the proposed 3D DetecTrack achieves significant improvements in both detection and tracking performances over baseline methods and achieves state-of-the-art performance among existing methods through collaboration between the detector and tracker.
The current state-of-the-art methods for video corpus moment retrieval (VCMR) often use similarity-based feature alignment approach for the sake of convenience and speed. However, late fusion methods like cosine similarity alignment are unable to make full use of the information from both query texts and videos. In this paper, we combine feature alignment with feature fusion to promote the performance on VCMR.
Merging datafiles containing information on overlapping sets of entities is a challenging task in the absence of unique identifiers, and is further complicated when some entities are duplicated in the datafiles. Most approaches to this problem have focused on linking two files assumed to be free of duplicates, or on detecting which records in a single file are duplicates. However, it is common in practice to encounter scenarios that fit somewhere in between or beyond these two settings. We propose a Bayesian approach for the general setting of multifile record linkage and duplicate detection. We use a novel partition representation to propose a structured prior for partitions that can incorporate prior information about the data collection processes of the datafiles in a flexible manner, and extend previous models for comparison data to accommodate the multifile setting. We also introduce a family of loss functions to derive Bayes estimates of partitions that allow uncertain portions of the partitions to be left unresolved. The performance of our proposed methodology is explored through extensive simulations. Code implementing the methodology is available at https://github.com/aleshing/multilink .
We introduce a deep neural network learning scheme to learn the B\"acklund transforms (BTs) of soliton evolution equations and an enhanced deep learning scheme for data-driven soliton equation discovery based on the known BTs, respectively. The first scheme takes advantage of some solution (or soliton equation) information to study the data-driven BT of sine-Gordon equation, and complex and real Miura transforms between the defocusing (focusing) mKdV equation and KdV equation, as well as the data-driven mKdV equation discovery via the Miura transforms. The second deep learning scheme uses the explicit/implicit BTs generating the higher-order solitons to train the data-driven discovery of mKdV and sine-Gordon equations, in which the high-order solution informations are more powerful for the enhanced leaning soliton equations with higher accurates.
This paper presents a novel graph-based kernel learning approach for connectome analysis. Specifically, we demonstrate how to leverage the naturally available structure within the graph representation to encode prior knowledge in the kernel. We first proposed a matrix factorization to directly extract structural features from natural symmetric graph representations of connectome data. We then used them to derive a structure-persevering graph kernel to be fed into the support vector machine. The proposed approach has the advantage of being clinically interpretable. Quantitative evaluations on challenging HIV disease classification (DTI- and fMRI-derived connectome data) and emotion recognition (EEG-derived connectome data) tasks demonstrate the superior performance of our proposed methods against the state-of-the-art. Results showed that relevant EEG-connectome information is primarily encoded in the alpha band during the emotion regulation task.
Image fusion methods and metrics for their evaluation have conventionally used pixel-based or low-level features. However, for many applications, the aim of image fusion is to effectively combine the semantic content of the input images. This paper proposes a novel system for the semantic combination of visual content using pre-trained CNN network architectures. Our proposed semantic fusion is initiated through the fusion of the top layer feature map outputs (for each input image)through gradient updating of the fused image input (so-called image optimisation). Simple "choose maximum" and "local majority" filter based fusion rules are utilised for feature map fusion. This provides a simple method to combine layer outputs and thus a unique framework to fuse single-channel and colour images within a decomposition pre-trained for classification and therefore aligned with semantic fusion. Furthermore, class activation mappings of each input image are used to combine semantic information at a higher level. The developed methods are able to give equivalent low-level fusion performance to state of the art methods while providing a unique architecture to combine semantic information from multiple images.
Video object detection is challenging in the presence of appearance deterioration in certain video frames. Therefore, it is a natural choice to aggregate temporal information from other frames of the same video into the current frame. However, RoI Align, as one of the most core procedures of video detectors, still remains extracting features from a single-frame feature map for proposals, making the extracted RoI features lack temporal information from videos. In this work, considering the features of the same object instance are highly similar among frames in a video, a novel Temporal RoI Align operator is proposed to extract features from other frames feature maps for current frame proposals by utilizing feature similarity. The proposed Temporal RoI Align operator can extract temporal information from the entire video for proposals. We integrate it into single-frame video detectors and other state-of-the-art video detectors, and conduct quantitative experiments to demonstrate that the proposed Temporal RoI Align operator can consistently and significantly boost the performance. Besides, the proposed Temporal RoI Align can also be applied into video instance segmentation. Codes are available at https://github.com/open-mmlab/mmtracking