Contrastive learning has demonstrated great effectiveness in representation learning especially for image classification tasks. However, there is still a shortage in the studies targeting regression tasks, and more specifically applications on hyperspectral data. In this paper, we propose a contrastive learning framework for the regression tasks for hyperspectral data. To this end, we provide a collection of transformations relevant for augmenting hyperspectral data, and investigate contrastive learning for regression. Experiments on synthetic and real hyperspectral datasets show that the proposed framework and transformations significantly improve the performance of regression models, achieving better scores than other state-of-the-art transformations.
Existing self-supervised methods in natural language processing (NLP), especially hierarchical text classification (HTC), mainly focus on self-supervised contrastive learning, extremely relying on human-designed augmentation rules to generate contrastive samples, which can potentially corrupt or distort the original information. In this paper, we tend to investigate the feasibility of a contrastive learning scheme in which the semantic and syntactic information inherent in the input sample is adequately reserved in the contrastive samples and fused during the learning process. Specifically, we propose an information lossless contrastive learning strategy for HTC, namely \textbf{H}ierarchy-aware \textbf{I}nformation \textbf{L}ossless contrastive \textbf{L}earning (HILL), which consists of a text encoder representing the input document, and a structure encoder directly generating the positive sample. The structure encoder takes the document embedding as input, extracts the essential syntactic information inherent in the label hierarchy with the principle of structural entropy minimization, and injects the syntactic information into the text representation via hierarchical representation learning. Experiments on three common datasets are conducted to verify the superiority of HILL.
Graph representation learning is a fundamental research issue in various domains of applications, of which the inductive learning problem is particularly challenging as it requires models to generalize to unseen graph structures during inference. In recent years, graph neural networks (GNNs) have emerged as powerful graph models for inductive learning tasks such as node classification, whereas they typically heavily rely on the annotated nodes under a fully supervised training setting. Compared with the GNN-based methods, variational graph auto-encoders (VGAEs) are known to be more generalizable to capture the internal structural information of graphs independent of node labels and have achieved prominent performance on multiple unsupervised learning tasks. However, so far there is still a lack of work focusing on leveraging the VGAE framework for inductive learning, due to the difficulties in training the model in a supervised manner and avoiding over-fitting the proximity information of graphs. To solve these problems and improve the model performance of VGAEs for inductive graph representation learning, in this work, we propose the Self-Label Augmented VGAE model. To leverage the label information for training, our model takes node labels as one-hot encoded inputs and then performs label reconstruction in model training. To overcome the scarcity problem of node labels for semi-supervised settings, we further propose the Self-Label Augmentation Method (SLAM), which uses pseudo labels generated by our model with a node-wise masking approach to enhance the label information. Experiments on benchmark inductive learning graph datasets verify that our proposed model archives promising results on node classification with particular superiority under semi-supervised learning settings.
Despite their black-box nature, deep learning models are extensively used in image-based drug discovery to extract feature vectors from single cells in microscopy images. To better understand how these networks perform representation learning, we employ visual explainability techniques (e.g., Grad-CAM). Our analyses reveal several mechanisms by which supervised models cheat, exploiting biologically irrelevant pixels when extracting morphological features from images, such as noise in the background. This raises doubts regarding the fidelity of learned single-cell representations and their relevance when investigating downstream biological questions. To address this misalignment between researcher expectations and machine behavior, we introduce Grad-CAMO, a novel single-cell interpretability score for supervised feature extractors. Grad-CAMO measures the proportion of a model's attention that is concentrated on the cell of interest versus the background. This metric can be assessed per-cell or averaged across a validation set, offering a tool to audit individual features vectors or guide the improved design of deep learning architectures. Importantly, Grad-CAMO seamlessly integrates into existing workflows, requiring no dataset or model modifications, and is compatible with both 2D and 3D Cell Painting data. Additional results are available at https://github.com/eigenvivek/Grad-CAMO.
We investigate a fundamental aspect of machine vision: the measurement of features, by revisiting clustering, one of the most classic approaches in machine learning and data analysis. Existing visual feature extractors, including ConvNets, ViTs, and MLPs, represent an image as rectangular regions. Though prevalent, such a grid-style paradigm is built upon engineering practice and lacks explicit modeling of data distribution. In this work, we propose feature extraction with clustering (FEC), a conceptually elegant yet surprisingly ad-hoc interpretable neural clustering framework, which views feature extraction as a process of selecting representatives from data and thus automatically captures the underlying data distribution. Given an image, FEC alternates between grouping pixels into individual clusters to abstract representatives and updating the deep features of pixels with current representatives. Such an iterative working mechanism is implemented in the form of several neural layers and the final representatives can be used for downstream tasks. The cluster assignments across layers, which can be viewed and inspected by humans, make the forward process of FEC fully transparent and empower it with promising ad-hoc interpretability. Extensive experiments on various visual recognition models and tasks verify the effectiveness, generality, and interpretability of FEC. We expect this work will provoke a rethink of the current de facto grid-style paradigm.
When benefiting graphic sketch representation with sketch drawing orders, recent studies have linked sketch patches as graph edges by drawing orders in accordance to a temporal-based nearest neighboring strategy. However, such constructed graph edges may be unreliable, since the contextual relationships between patches may be inconsistent with the sequential positions in drawing orders, due to variants of sketch drawings. In this paper, we propose a variant-drawing-protected method by equipping sketch patches with context-aware positional encoding (PE) to make better use of drawing orders for sketch learning. We introduce a sinusoidal absolute PE to embed the sequential positions in drawing orders, and a learnable relative PE to encode the unseen contextual relationships between patches. Both types of PEs never attend the construction of graph edges, but are injected into graph nodes to cooperate with the visual patterns captured from patches. After linking nodes by semantic proximity, during message aggregation via graph convolutional networks, each node receives both semantic features from patches and contextual information from PEs from its neighbors, which equips local patch patterns with global contextual information, further obtaining drawing-order-enhanced sketch representations. Experimental results indicate that our method significantly improves sketch healing and controllable sketch synthesis. The source codes could be found at https://github.com/SCZang/DC-gra2seq.
Graph representation learning has become a crucial task in machine learning and data mining due to its potential for modeling complex structures such as social networks, chemical compounds, and biological systems. Spiking neural networks (SNNs) have recently emerged as a promising alternative to traditional neural networks for graph learning tasks, benefiting from their ability to efficiently encode and process temporal and spatial information. In this paper, we propose a novel approach that integrates attention mechanisms with SNNs to improve graph representation learning. Specifically, we introduce an attention mechanism for SNN that can selectively focus on important nodes and corresponding features in a graph during the learning process. We evaluate our proposed method on several benchmark datasets and show that it achieves comparable performance compared to existing graph learning techniques.
Graph Transformers (GTs) with powerful representation learning ability make a huge success in wide range of graph tasks. However, the costs behind outstanding performances of GTs are higher energy consumption and computational overhead. The complex structure and quadratic complexity during attention calculation in vanilla transformer seriously hinder its scalability on the large-scale graph data. Though existing methods have made strides in simplifying combinations among blocks or attention-learning paradigm to improve GTs' efficiency, a series of energy-saving solutions originated from biologically plausible structures are rarely taken into consideration when constructing GT framework. To this end, we propose a new spiking-based graph transformer (SGHormer). It turns full-precision embeddings into sparse and binarized spikes to reduce memory and computational costs. The spiking graph self-attention and spiking rectify blocks in SGHormer explicitly capture global structure information and recover the expressive power of spiking embeddings, respectively. In experiments, SGHormer achieves comparable performances to other full-precision GTs with extremely low computational energy consumption. The results show that SGHomer makes a remarkable progress in the field of low-energy GTs.
Multilingual pretraining and fine-tuning have remarkably succeeded in various natural language processing tasks. Transferring representations from one language to another is especially crucial for cross-lingual learning. One can expect machine translation objectives to be well suited to fostering such capabilities, as they involve the explicit alignment of semantically equivalent sentences from different languages. This paper investigates the potential benefits of employing machine translation as a continued training objective to enhance language representation learning, bridging multilingual pretraining and cross-lingual applications. We study this question through two lenses: a quantitative evaluation of the performance of existing models and an analysis of their latent representations. Our results show that, contrary to expectations, machine translation as the continued training fails to enhance cross-lingual representation learning in multiple cross-lingual natural language understanding tasks. We conclude that explicit sentence-level alignment in the cross-lingual scenario is detrimental to cross-lingual transfer pretraining, which has important implications for future cross-lingual transfer studies. We furthermore provide evidence through similarity measures and investigation of parameters that this lack of positive influence is due to output separability -- which we argue is of use for machine translation but detrimental elsewhere.
Automatic prostate segmentation is an important step in computer-aided diagnosis of prostate cancer and treatment planning. Existing methods of prostate segmentation are based on deep learning models which have a large size and lack of transparency which is essential for physicians. In this paper, a new data-driven 3D prostate segmentation method on MRI is proposed, named PSHop. Different from deep learning based methods, the core methodology of PSHop is a feed-forward encoder-decoder system based on successive subspace learning (SSL). It consists of two modules: 1) encoder: fine to coarse unsupervised representation learning with cascaded VoxelHop units, 2) decoder: coarse to fine segmentation prediction with voxel-wise classification and local refinement. Experiments are conducted on the publicly available ISBI-2013 dataset, as well as on a larger private one. Experimental analysis shows that our proposed PSHop is effective, robust and lightweight in the tasks of prostate gland and zonal segmentation, achieving a Dice Similarity Coefficient (DSC) of 0.873 for the gland segmentation task. PSHop achieves a competitive performance comparatively to other deep learning methods, while keeping the model size and inference complexity an order of magnitude smaller.
Despite living in a multi-sensory world, most AI models are limited to textual and visual understanding of human motion and behavior. In fact, full situational awareness of human motion could best be understood through a combination of sensors. In this survey we investigate how knowledge can be transferred and utilized amongst modalities for Human Activity/Action Recognition (HAR), i.e. cross-modality transfer learning. We motivate the importance and potential of IMU data and its applicability in cross-modality learning as well as the importance of studying the HAR problem. We categorize HAR related tasks by time and abstractness and then compare various types of multimodal HAR datasets. We also distinguish and expound on many related but inconsistently used terms in the literature, such as transfer learning, domain adaptation, representation learning, sensor fusion, and multimodal learning, and describe how cross-modal learning fits with all these concepts. We then review the literature in IMU-based cross-modal transfer for HAR. The two main approaches for cross-modal transfer are instance-based transfer, where instances of one modality are mapped to another (e.g. knowledge is transferred in the input space), or feature-based transfer, where the model relates the modalities in an intermediate latent space (e.g. knowledge is transferred in the feature space). Finally, we discuss future research directions and applications in cross-modal HAR.
Causal representation learning seeks to uncover latent, high-level causal representations from low-level observed data. It is particularly good at predictions under unseen distribution shifts, because these shifts can generally be interpreted as consequences of interventions. Hence leveraging {seen} distribution shifts becomes a natural strategy to help identifying causal representations, which in turn benefits predictions where distributions are previously {unseen}. Determining the types (or conditions) of such distribution shifts that do contribute to the identifiability of causal representations is critical. This work establishes a {sufficient} and {necessary} condition characterizing the types of distribution shifts for identifiability in the context of latent additive noise models. Furthermore, we present partial identifiability results when only a portion of distribution shifts meets the condition. In addition, we extend our findings to latent post-nonlinear causal models. We translate our findings into a practical algorithm, allowing for the acquisition of reliable latent causal representations. Our algorithm, guided by our underlying theory, has demonstrated outstanding performance across a diverse range of synthetic and real-world datasets. The empirical observations align closely with the theoretical findings, affirming the robustness and effectiveness of our approach.
Graph Neural Networks (GNNs) have emerged as the most powerful weapon for various graph tasks due to the message-passing mechanism's great local information aggregation ability. However, over-smoothing has always hindered GNNs from going deeper and capturing multi-hop neighbors. Unlike GNNs, Transformers can model global information and multi-hop interactions via multi-head self-attention and a proper Transformer structure can show more immunity to the over-smoothing problem. So, can we propose a novel framework to combine GNN and Transformer, integrating both GNN's local information aggregation and Transformer's global information modeling ability to eliminate the over-smoothing problem? To realize this, this paper proposes a collaborative learning scheme for GNN-Transformer and constructs GTC architecture. GTC leverages the GNN and Transformer branch to encode node information from different views respectively, and establishes contrastive learning tasks based on the encoded cross-view information to realize self-supervised heterogeneous graph representation. For the Transformer branch, we propose Metapath-aware Hop2Token and CG-Hetphormer, which can cooperate with GNN to attentively encode neighborhood information from different levels. As far as we know, this is the first attempt in the field of graph representation learning to utilize both GNN and Transformer to collaboratively capture different view information and conduct cross-view contrastive learning. The experiments on real datasets show that GTC exhibits superior performance compared with state-of-the-art methods. Codes can be available at https://github.com/PHD-lanyu/GTC.
Alzheimer's disease (AD) is an incurable neurodegenerative condition leading to cognitive and functional deterioration. Given the lack of a cure, prompt and precise AD diagnosis is vital, a complex process dependent on multiple factors and multi-modal data. While successful efforts have been made to integrate multi-modal representation learning into medical datasets, scant attention has been given to 3D medical images. In this paper, we propose Contrastive Masked Vim Autoencoder (CMViM), the first efficient representation learning method tailored for 3D multi-modal data. Our proposed framework is built on a masked Vim autoencoder to learn a unified multi-modal representation and long-dependencies contained in 3D medical images. We also introduce an intra-modal contrastive learning module to enhance the capability of the multi-modal Vim encoder for modeling the discriminative features in the same modality, and an inter-modal contrastive learning module to alleviate misaligned representation among modalities. Our framework consists of two main steps: 1) incorporate the Vision Mamba (Vim) into the mask autoencoder to reconstruct 3D masked multi-modal data efficiently. 2) align the multi-modal representations with contrastive learning mechanisms from both intra-modal and inter-modal aspects. Our framework is pre-trained and validated ADNI2 dataset and validated on the downstream task for AD classification. The proposed CMViM yields 2.7\% AUC performance improvement compared with other state-of-the-art methods.
Self-supervised learning has emerged as a powerful tool for pretraining deep networks on unlabeled data, prior to transfer learning of target tasks with limited annotation. The relevance between the pretraining pretext and target tasks is crucial to the success of transfer learning. Various pretext tasks have been proposed to utilize properties of medical image data (e.g., three dimensionality), which are more relevant to medical image analysis than generic ones for natural images. However, previous work rarely paid attention to data with anatomy-oriented imaging planes, e.g., standard cardiac magnetic resonance imaging views. As these imaging planes are defined according to the anatomy of the imaged organ, pretext tasks effectively exploiting this information can pretrain the networks to gain knowledge on the organ of interest. In this work, we propose two complementary pretext tasks for this group of medical image data based on the spatial relationship of the imaging planes. The first is to learn the relative orientation between the imaging planes and implemented as regressing their intersecting lines. The second exploits parallel imaging planes to regress their relative slice locations within a stack. Both pretext tasks are conceptually straightforward and easy to implement, and can be combined in multitask learning for better representation learning. Thorough experiments on two anatomical structures (heart and knee) and representative target tasks (semantic segmentation and classification) demonstrate that the proposed pretext tasks are effective in pretraining deep networks for remarkably boosted performance on the target tasks, and superior to other recent approaches.
Graph neural networks have achieved remarkable success in learning graph representations, especially graph Transformer, which has recently shown superior performance on various graph mining tasks. However, graph Transformer generally treats nodes as tokens, which results in quadratic complexity regarding the number of nodes during self-attention computation. The graph MLP Mixer addresses this challenge by using the efficient MLP Mixer technique from computer vision. However, the time-consuming process of extracting graph tokens limits its performance. In this paper, we present a novel architecture named ChebMixer, a newly graph MLP Mixer that uses fast Chebyshev polynomials-based spectral filtering to extract a sequence of tokens. Firstly, we produce multiscale representations of graph nodes via fast Chebyshev polynomial-based spectral filtering. Next, we consider each node's multiscale representations as a sequence of tokens and refine the node representation with an effective MLP Mixer. Finally, we aggregate the multiscale representations of nodes through Chebyshev interpolation. Owing to the powerful representation capabilities and fast computational properties of MLP Mixer, we can quickly extract more informative node representations to improve the performance of downstream tasks. The experimental results prove our significant improvements in a variety of scenarios ranging from graph node classification to medical image segmentation.
This paper focuses on understanding the predominant video creation pipeline, i.e., compositional video editing with six main types of editing components, including video effects, animation, transition, filter, sticker, and text. In contrast to existing visual representation learning of visual materials (i.e., images/videos), we aim to learn visual representations of editing actions/components that are generally applied on raw materials. We start by proposing the first large-scale dataset for editing components of video creation, which covers about $3,094$ editing components with $618,800$ videos. Each video in our dataset is rendered by various image/video materials with a single editing component, which supports atomic visual understanding of different editing components. It can also benefit several downstream tasks, e.g., editing component recommendation, editing component recognition/retrieval, etc. Existing visual representation methods perform poorly because it is difficult to disentangle the visual appearance of editing components from raw materials. To that end, we benchmark popular alternative solutions and propose a novel method that learns to attend to the appearance of editing components regardless of raw materials. Our method achieves favorable results on editing component retrieval/recognition compared to the alternative solutions. A user study is also conducted to show that our representations cluster visually similar editing components better than other alternatives. Furthermore, our learned representations used to transition recommendation tasks achieve state-of-the-art results on the AutoTransition dataset. The code and dataset will be released for academic use.
In this paper, we propose a new model to learn Adaptive Kernel-based Representations (AKBR) for graph classification. Unlike state-of-the-art R-convolution graph kernels that are defined by merely counting any pair of isomorphic substructures between graphs and cannot provide an end-to-end learning mechanism for the classifier, the proposed AKBR approach aims to define an end-to-end representation learning model to construct an adaptive kernel matrix for graphs. To this end, we commence by leveraging a novel feature-channel attention mechanism to capture the interdependencies between different substructure invariants of original graphs. The proposed AKBR model can thus effectively identify the structural importance of different substructures, and compute the R-convolution kernel between pairwise graphs associated with the more significant substructures specified by their structural attentions. Since each row of the resulting kernel matrix can be theoretically seen as the embedding vector of a sample graph, the proposed AKBR model is able to directly employ the resulting kernel matrix as the graph feature matrix and input it into the classifier for classification (i.e., the SoftMax layer), naturally providing an end-to-end learning architecture between the kernel computation as well as the classifier. Experimental results show that the proposed AKBR model outperforms existing state-of-the-art graph kernels and deep learning methods on standard graph benchmarks.
While intrusion detection systems (IDSs) benefit from the diversity and generalization of IoT data features, the data diversity (e.g., the heterogeneity and high dimensions of data) also makes it difficult to train effective machine learning models in IoT IDSs. This also leads to potentially redundant/noisy features that may decrease the accuracy of the detection engine in IDSs. This paper first introduces a novel neural network architecture called Multiple-Input Auto-Encoder (MIAE). MIAE consists of multiple sub-encoders that can process inputs from different sources with different characteristics. The MIAE model is trained in an unsupervised learning mode to transform the heterogeneous inputs into lower-dimensional representation, which helps classifiers distinguish between normal behaviour and different types of attacks. To distil and retain more relevant features but remove less important/redundant ones during the training process, we further design and embed a feature selection layer right after the representation layer of MIAE resulting in a new model called MIAEFS. This layer learns the importance of features in the representation vector, facilitating the selection of informative features from the representation vector. The results on three IDS datasets, i.e., NSLKDD, UNSW-NB15, and IDS2017, show the superior performance of MIAE and MIAEFS compared to other methods, e.g., conventional classifiers, dimensionality reduction models, unsupervised representation learning methods with different input dimensions, and unsupervised feature selection models. Moreover, MIAE and MIAEFS combined with the Random Forest (RF) classifier achieve accuracy of 96.5% in detecting sophisticated attacks, e.g., Slowloris. The average running time for detecting an attack sample using RF with the representation of MIAE and MIAEFS is approximate 1.7E-6 seconds, whilst the model size is lower than 1 MB.
Representation learning (RL) methods for cyberattack detection face the diversity and sophistication of attack data, leading to the issue of mixed representations of different classes, particularly as the number of classes increases. To address this, the paper proposes a novel deep learning architecture/model called the Twin Auto-Encoder (TAE). TAE first maps the input data into latent space and then deterministically shifts data samples of different classes further apart to create separable data representations, referred to as representation targets. TAE's decoder then projects the input data into these representation targets. After training, TAE's decoder extracts data representations. TAE's representation target serves as a novel dynamic codeword, which refers to the vector that represents a specific class. This vector is updated after each training epoch for every data sample, in contrast to the conventional fixed codeword that does not incorporate information from the input data. We conduct extensive experiments on diverse cybersecurity datasets, including seven IoT botnet datasets, two network IDS datasets, three malware datasets, one cloud DDoS dataset, and ten artificial datasets as the number of classes increases. TAE boosts accuracy and F-score in attack detection by around 2% compared to state-of-the-art models, achieving up to 96.1% average accuracy in IoT attack detection. Additionally, TAE is well-suited for cybersecurity applications and potentially for IoT systems, with a model size of approximately 1 MB and an average running time of around 2.6E-07 seconds for extracting a data sample.