Single Embedding Research Articles

A comprehensive analysis of static word embeddings for Turkish

Word embeddings are fixed-length, dense and distributed word representations that are used in natural language processing (NLP) applications. There are basically two types of word embedding models which are non-contextual (static) models and contextual models. The former method generates a single embedding for a word regardless of its context, while the latter method produces distinct embeddings for a word based on the specific contexts in which it appears. There are plenty of works that compare contextual and non-contextual embedding models within their respective groups in different languages. However, the number of studies that compare the models in these two groups with each other is very few and there is no such study in Turkish. This process necessitates converting contextual embeddings into static embeddings. In this paper, we compare and evaluate the performance of several contextual and non-contextual models in both intrinsic and extrinsic evaluation settings for Turkish. We make a fine-grained comparison by analyzing the syntactic and semantic capabilities of the models separately. The results of the analyses provide insights about the suitability of different embedding models in different types of NLP tasks. We also build a Turkish word embedding repository comprising the embedding models used in this work, which may serve as a valuable resource for researchers and practitioners in the field of Turkish NLP. We make the word embeddings, scripts, and evaluation datasets publicly available.

Decoupled Textual Embeddings for Customized Image Generation

Customized text-to-image generation, which aims to learn user-specified concepts with a few images, has drawn significant attention recently. However, existing methods usually suffer from overfitting issues and entangle the subject-unrelated information (e.g., background and pose) with the learned concept, limiting the potential to compose concept into new scenes. To address these issues, we propose the DETEX, a novel approach that learns the disentangled concept embedding for flexible customized text-to-image generation. Unlike conventional methods that learn a single concept embedding from the given images, our DETEX represents each image using multiple word embeddings during training, i.e., a learnable image-shared subject embedding and several image-specific subject-unrelated embeddings. To decouple irrelevant attributes (i.e., background and pose) from the subject embedding, we further present several attribute mappers that encode each image as several image-specific subject-unrelated embeddings. To encourage these unrelated embeddings to capture the irrelevant information, we incorporate them with corresponding attribute words and propose a joint training strategy to facilitate the disentanglement. During inference, we only use the subject embedding for image generation, while selectively using image-specific embeddings to retain image-specified attributes. Extensive experiments demonstrate that the subject embedding obtained by our method can faithfully represent the target concept, while showing superior editability compared to the state-of-the-art methods. Our code will be available at https://github.com/PrototypeNx/DETEX.

Sync-NeRF: Generalizing Dynamic NeRFs to Unsynchronized Videos

Recent advancements in 4D scene reconstruction using neural radiance fields (NeRF) have demonstrated the ability to represent dynamic scenes from multi-view videos. However, they fail to reconstruct the dynamic scenes and struggle to fit even the training views in unsynchronized settings. It happens because they employ a single latent embedding for a frame while the multi-view images at the same frame were actually captured at different moments. To address this limitation, we introduce time offsets for individual unsynchronized videos and jointly optimize the offsets with NeRF. By design, our method is applicable for various baselines and improves them with large margins. Furthermore, finding the offsets always works as synchronizing the videos without manual effort. Experiments are conducted on the common Plenoptic Video Dataset and a newly built Unsynchronized Dynamic Blender Dataset to verify the performance of our method. Project page: https://seoha-kim.github.io/sync-nerf

A multi-embedding domain video steganography algorithm based on TU partitioning and intra prediction mode

With High Efficiency Video Coding (HEVC) becoming a popular video coding standard in the world, combing the HEVC with data hiding methods is a complex task. In this paper, a multi-embedded video steganography algorithm based on Intra Prediction Mode (IPM) and Transform Unit (TU) partition is proposed. By analyzing the coding characteristics of the TU partition and the IPM, we select embedded regions that are relatively independent as much as possible. Then, suitable distortion functions are designed for different embedding domains. Using the minimum distortion framework of the STC model to embed messages. The experimental results indicate that the proposed method achieves a maximum embedding capacity 1.5 times larger than that of traditional single domain embedding methods, while maintaining good visual quality.

Deep Multirepresentation Learning for Data Clustering.

Deep clustering incorporates embedding into clustering in order to find a lower-dimensional space suitable for clustering tasks. Conventional deep clustering methods aim to obtain a single global embedding subspace (aka latent space) for all the data clusters. In contrast, in this article, we propose a deep multirepresentation learning (DML) framework for data clustering whereby each difficult-to-cluster data group is associated with its own distinct optimized latent space and all the easy-to-cluster data groups are associated with a general common latent space. Autoencoders (AEs) are employed for generating cluster-specific and general latent spaces. To specialize each AE in its associated data cluster(s), we propose a novel and effective loss function which consists of weighted reconstruction and clustering losses of the data points, where higher weights are assigned to the samples more probable to belong to the corresponding cluster(s). Experimental results on benchmark datasets demonstrate that the proposed DML framework and loss function outperform state-of-the-art clustering approaches. In addition, the results show that the DML method significantly outperforms the SOTA on imbalanced datasets as a result of assigning an individual latent space to the difficult clusters.

Keyword-Based Diverse Image Retrieval With Variational Multiple Instance Graph.

The task of cross-modal image retrieval has recently attracted considerable research attention. In real-world scenarios, keyword-based queries issued by users are usually short and have broad semantics. Therefore, semantic diversity is as important as retrieval accuracy in such user-oriented services, which improves user experience. However, most typical cross-modal image retrieval methods based on single point query embedding inevitably result in low semantic diversity, while existing diverse retrieval approaches frequently lead to low accuracy due to a lack of cross-modal understanding. To address this challenge, we introduce an end-to-end solution termed variational multiple instance graph (VMIG), in which a continuous semantic space is learned to capture diverse query semantics, and the retrieval task is formulated as a multiple instance learning problems to connect diverse features across modalities. Specifically, a query-guided variational autoencoder is employed to model the continuous semantic space instead of learning a single-point embedding. Afterward, multiple instances of the image and query are obtained by sampling in the continuous semantic space and applying multihead attention, respectively. Thereafter, an instance graph is constructed to remove noisy instances and align cross-modal semantics. Finally, heterogeneous modalities are robustly fused under multiple losses. Extensive experiments on two real-world datasets have well verified the effectiveness of our proposed solution in both retrieval accuracy and semantic diversity.

OneSpace: Detecting cross-language clones by learning a common embedding space

Identifying clone code fragments across different languages can enhance the productivity of software developers in several ways. However, the clone detection task is often studied in the context of a single language and less explored for code snippets spanning different languages. In this paper, we present OneSpace, a new cross-language clone detection approach. OneSpace projects different programming languages to the same embedding space using both code and API data. OneSpace, hence, leverages a Siamese Network to infer the similarity of the embedded programs. We evaluate OneSpace by detecting clones across three language pairs; JAVA-Python, Java-C++ and Java-C. We compared OneSpace with the other state-of-art techniques, SupLearn and CLCDSA. In our evaluation, OneSpace provided higher effectiveness than the state of the art. Our ablation study validated some of our intuitions in designing OneSpace, particularly that using a single embedding space (as opposed to separate ones) provides higher effectiveness. Additionally, we designed a variant of OneSpace that uses Word-Mover-Distance Algorithm and provides lower effectiveness, but is much more efficient. We also found that OneSpace provides higher effectiveness than the state of the art, even for: complex implementations, single-method implementations, varying ratios of positive to negative clones in training, varying amounts of training data, and for additional programming languages.

Promoted degradation of tetracycline by FeNi alloys embedding in biochar activating peroxydisulfate: Performance and mechanism

Bimetallic FeNi alloys embedding into biochar (FeNi-BC) catalysts with different Fe/Ni molar ratio were fabricated using rose petal as biomaterials. The as-fabricated FeNi-BC was employed to activate peroxydisulfate (PDS) toward tetracycline (TC) degradation. The effects of Fe/Ni atomic ratio on the structure feature of FeNi-BC and PDS activation efficiency were explored in detail. Results show that the optimum Fe0.8Ni0.2-BC prepared with Fe/Ni molar being 8:2 exhibited the best PDS activation efficiency toward TC degradation, with 97.01% of TC being removed in 15 min. Combining with the electron paramagnetic resonance (EPR), reactive oxygen species quenching, and electrochemical tests, the direct electron transfer (DET) and singlet oxygen accounted for TC degradation in the Fe0.8Ni0.2-BC activating PDS system. Because of that, the Fe0.8Ni0.2-BC had satisfactory resistance to inorganic anions and natural organic matters (NOMs). Based on the structure-activity relationship, it was found that more Fe0 and pyrrolic N formation in FeNi alloys than single Fe or Ni embedding in BC facilitated in more PDS adsorption, thus more PDS* complex formation on the surface of catalysts to boost the TC degradation through DET mechanism. Moreover, Fe0.8Ni0.2-BC possessed trace Fe and Ni leaching, satisfactory stability, and easy recycling via magnetic separation. This study not only offers new insights for discriminating the catalytic active sites of biochar embedded bimetallic alloys catalysts activating PDS but also developed cost-effective and environment friendly catalysts for high-efficiently remove antibiotics via PS-AOPs.

A Multimodal Single Cell Atlas of Adult Murine Haematopoiesis Reveals Nine Differentiation Trajectories and Epigenetic Regulatory Programs

Single cell genomics has revolutionised our understanding of differentiating systems and led to a reinterpretation of haematopoiesis, moving from a step-wise hierarchical tree model towards more continous differentiation landscapes. Recent single cell RNA sequencing (scRNA-seq) studies of Lineage- ckit+ (LK) mouse bone marrow progenitors have defined a highly granular landscape of differentiation from multipotent stem cells to committed progenitors in 8 different lineages and have served as a valuable reference landscape for comparisons with perturbation states. While these landscapes have permitted exploration and interrogation of early haematopoietic progenitors, as yet, no map of mouse total bone marrow haematopoiesis exists at single cell resolution, limiting efforts to fully understand the gene programs which specify and the cell surface phenotypes which identify maturing haematopoietic lineages nor the connection between early bifurcation decisions in progenitor cells and mature cellular outputs. To answer these questions, we performed droplet-based scRNA-seq and cell surface proteomics on whole bone marrow from 4 male and 4 female 12-week old C57Bl6 mice, using a panel of 138 antibodies against cell surface antigens. To increase representation of rare progenitor populations, we complemented total bone marrow mononuclear cells with c-kit enrichment and FACS sorted LK and LK sca1-positive (LSK) populations. After quality control filtering the single cell transcriptomes and proteomes, over 198,000 single cells were analysed separately before multimodal integration into a single UMAP embedding. Addition of cell surface proteome information to the scRNA-seq data through multimodal integration produced a representation with greater resolution of haematopoietic cell types, including improved resolution of T-cell subtypes and separation of a rare population of eosinophil progenitors which were previously indistinguishable from basophil progenitors. This integrated landscape of mouse haematopoiesis was then clustered and annotated though an iterative process of label transfer and manual annotation into 41 cell type states (Figure 1). Using cell fate analysis, which utilises transcriptional similarity and pseudotemporal ordering to estimate the probability of a single cell's commitment to defined terminal states, we were able to resolve cell fate probabilities towards nine haematopoietic lineages (erythroid, neutrophil, megakaryocyte, lymphoid, monocyte, dendritic cell (DC), plasmacytoid dendritic cell (pDC), basophil and eosinophil (Figure 2)). From these probabilities we defined nine trajectories from the earliest uncommitted progenitors to terminally differentiated populations, including a pDC trajectory involving both lymphoid and common DC progenitors. Analysis of trajectory-based gene expression trends provides a framework for discovery of lineage drivers as well as shared gene expression programs across multiple lineages. Leveraging cell surface proteome data permits the use of ‘in-silico-FACS’ to describe phenotypes of cell populations which exist at cell fate branch-points, for further in-vitro characterisation. To demonstrate the utility of our atlas as a reference, we projected perturbation datasets from 7 pre-leukaemic mouse models onto the atlas, allowing quantitation of mutation-associated cellular and tissue-scale alterations. Finally, we have leveraged the transcription-start site enrichment of 5' scRNA-seq to compute enhancer RNA (eRNA) expression in our atlas, allowing the identification of dynamic usage of specific enhancer elements during haematopoietic specification and thus providing a novel framework for correlation of eRNA expression (as a proxy for active distal chromatin regions) with gene expression trends along specific haematopoietic lineages.

HCURec: Hierarchical candidate-aware user modeling for news recommendation

Matching candidate news with user interests is critical for news recommendation. Current studies on news recommendation mainly model a single user interest embedding from the user’s clicked news. One of the major challenges that affects the recommendation is to match candidate news with the user’s multi-field and multi-grained interests. To confront this difficulty, we investigate the multi-grained correlation of user interests and candidate news to obtain their matching features. We propose a hierarchical candidate-aware user modeling framework for news recommendation that matches users’ multi-field and multi-grained interests with candidate news. The framework incorporates candidate news into the modeling of user interests at different levels, namely subcategory-level, category-level and global-level, which learn fine-grained, coarse-grained and overall user interests, respectively. The user interest is finally hierarchically matched at different levels with candidate news to achieve accurate targeting. A collection of experiments were carried out on four real large-scale datasets, and the experimental outcomes demonstrate the supremacy of the proposed method over the existing state-of-the-art methods owing to its highly effective and efficient performance of news recommendation.

Data-driven Morphological Feature Perception of Single Neuron with Graph Neural Network.

Clarifying the morphological characteristics of neurons can promote the understanding of brain function. However, traditional morphometrics fail to capture the modeling of each point in reconstructed neurons, leading to limited ability to distinguish massive nerve fibers and restricted application scenarios. To address these challenges, we propose MorphoGNN, a single neuron morphological embedding based on a graph neural network in this study. MorphoGNN learns the point-level structure information of reconstructed nerve fibers by considering their nearest neighbors on each hidden layer. This enables MorphoGNN to capture the lower-dimensional representation of a single neuron through an end-to-end model. In order to meet the requirements of various tasks, both supervised and self-supervised training strategies are designed to learn the characteristics that fit artificial semantics or the morphological patterns of neurons, respectively. We quantitatively compare our embeddings with other features in neuron classification and retrieval tasks and demonstrate cutting-edge performance. Additionally, we introduce our embeddings to the task of reconstruction quality classification and neuron clustering, where they can help detect reconstruction errors and obtain similar subtyping results to existing work. Furthermore, our method can be handily combined with other modal features, such as microscopic image features and traditional morphometrics. Ablation and robustness tests are also conducted to analyze the impact of several network components and low-quality reconstructed neurons on the performance of our method. The code is available at https://github.com/fun0515/MorphoGNN.

Prompt-based and weak-modality enhanced multimodal recommendation

Beyond conventional recommendation systems that rely merely on user-item interaction data, multimodal recommendation systems additionally exploit the item multimodal data for boosting the recommendation performance. In this research line, late fusion-based approaches that first predict user ratings for each item modality independently and then merge these predictions for a final user rating have made significant advancements. Nevertheless, these methods still have the following two issues: (1) they utilize individual user embeddings to model user interest in different modalities, while overlooking the underlying relationship among modalities and significantly increasing the memory costs; and (2) they overlook the unreliable interest learned from certain modality, thus hindering the accurate final rating learning. To address these issues, we propose a prompt-based and weak-modality enhanced multimodal recommendation framework. It consists of two key components: (1) multimodal prompted user interest learning that adopts a single user embedding with different modality prompts to model different modality-specific user interests, and (2) weak-modality enhanced training that enhances the user interest learning in modalities where the predictions are less unreliable, ensuring well-balanced learning across all modalities. Extensive experiments on Amazon datasets have demonstrated the effectiveness of the proposed framework. The two components deployed onto existing methods help to make them more effective and efficient.

EpilepsyNet: Novel automated detection of epilepsy using transformer model with EEG signals from 121 patient population

BackgroundEpilepsy is one of the most common neurological conditions globally, and the fourth most common in the United States. Recurrent non-provoked seizures characterize it and have huge impacts on the quality of life and financial impacts for affected individuals. A rapid and accurate diagnosis is essential in order to instigate and monitor optimal treatments. There is also a compelling need for the accurate interpretation of epilepsy due to the current scarcity in neurologist diagnosticians and a global inequity in access and outcomes. Furthermore, the existing clinical and traditional machine learning diagnostic methods exhibit limitations, warranting the need to create an automated system using deep learning model for epilepsy detection and monitoring using a huge database. MethodThe EEG signals from 35 channels were used to train the deep learning-based transformer model named (EpilepsyNet). For each training iteration, 1-min-long data were randomly sampled from each participant. Thereafter, each 5-s epoch was mapped to a matrix using the Pearson Correlation Coefficient (PCC), such that the bottom part of the triangle was discarded and only the upper triangle of the matrix was vectorized as input data. PCC is a reliable method used to measure the statistical relationship between two variables. Based on the 5 s of data, single embedding was performed thereafter to generate a 1-dimensional array of signals. In the final stage, a positional encoding with learnable parameters was added to each correlation coefficient's embedding before being fed to the developed EpilepsyNet as input data to epilepsy EEG signals. The ten-fold cross-validation technique was used to generate the model. ResultsOur transformer-based model (EpilepsyNet) yielded high classification accuracy, sensitivity, specificity and positive predictive values of 85%, 82%, 87%, and 82%, respectively. ConclusionThe proposed method is both accurate and robust since ten-fold cross-validation was employed to evaluate the performance of the model. Compared to the deep models used in existing studies for epilepsy diagnosis, our proposed method is simple and less computationally intensive. This is the earliest study to have uniquely employed the positional encoding with learnable parameters to each correlation coefficient's embedding together with the deep transformer model, using a huge database of 121 participants for epilepsy detection. With the training and validation of the model using a larger dataset, the same study approach can be extended for the detection of other neurological conditions, with a transformative impact on neurological diagnostics worldwide.

Botnet DGA Domain Name Classification Using Transformer Network with Hybrid Embedding

One of the severest threats to cyber security is botnet, which typically uses domain names generated by Domain Generation Algorithms (DGAs) to communicate with their Command and Control (C&C) infrastructure. DGA detection and classification play an important role of assisting cyber security researchers to detect botnet C&C servers. However, many of the existing DGA detection models only focus on single scale word embedding method, and very few models are specially designed to extract more effective features for DGA detection from multiple scales word embedding. To alleviate above questions, first we propose a hybrid word embedding method, which combines character level embedding and bigram level embedding to make full use of the domain names information, and then, we design a deep neural network with hybrid embedding method to distinguish DGA domains from known legitimate domains. Finally, we evaluate our hybrid embedding method and the proposed model on ONIST dataset and compare our methods with several state-of-the-art DGA classification methods.

Intraoperative cortical localization of music and language reveals signatures of structural complexity in posterior temporal cortex

Language and music involve the productive combination of basic units into structures. It remains unclear whether brain regions sensitive to linguistic and musical structure are co-localized. We report an intraoperative awake craniotomy in which a left-hemispheric language-dominant professional musician underwent cortical stimulation mapping (CSM) and electrocorticography of music and language perception and production during repetition tasks. Musical sequences were melodic or amelodic, and differed in algorithmic compressibility (Lempel-Ziv complexity). Auditory recordings of sentences differed in syntactic complexity (single vs. multiple phrasal embeddings). CSM of posterior superior temporal gyrus (pSTG) disrupted music perception and production, along with speech production. pSTG and posterior middle temporal gyrus (pMTG) activated for language and music (broadband gamma; 70-150Hz). pMTG activity was modulated by musical complexity, while pSTG activity was modulated by syntactic complexity. This points to shared resources for music and language comprehension, but distinct neural signatures for the processing of domain-specific structural features.

VIGA: A variational graph autoencoder model to infer user interest representations for recommendation

Learning representations of both user interests and item characteristics is essentially important for recommendation tasks. Although graph neural network-based methods have significant advantages, there are still two inherent limitations: (i) users or items are often modeled as single embedding points in the vector space, and (ii) these methods usually focus more on the graph structure of user–item interactions, rather than the attribute-related information and useful attribute interactions. These limitations make it difficult to model complex user interests and learn high-quality representations, thus fail to obtain high recommendation accuracy. To overcome these limitations, we propose a variational inference-based graph autoencoder (VIGA) model to explore a multivariate distribution over latent representations for recommendation. In brief, to identify both of the complex users' interests and their interactive behaviors, VIGA first encodes users and items into latent variables, and subjects them to a multivariate Gaussian distribution via a graph autoencoder framework based on Bayesian variational inference. Then a feature fusion layer is established for the exchange and integration of both the topological structure and nodes attributes in the user–item interaction graph. Experiments on three real-world datasets show that the proposed VIGA method significantly outperforms state-of-the-art recommendation methods on recommendation accuracy.

CSGVD: A deep learning approach combining sequence and graph embedding for source code vulnerability detection

In order to secure software, it is critical to detect potential vulnerabilities. The performance of traditional static vulnerability detection methods is limited by predefined rules, which rely heavily on the expertise of developers. Existing deep learning-based vulnerability detection models usually use only a single sequence or graph embedding approach to extract vulnerability features. Sequence embedding-based models ignore the structured information inherent in the code, and graph embedding-based models lack effective node and graph embedding methods. As a result, we propose a novel deep learning-based approach, CSGVD (Combining Sequence and Graph embedding for Vulnerability Detection), which considers function-level vulnerability detection as a graph binary classification task. Firstly, we propose a PE-BL module, which inherits and enhances the knowledge from the pre-trained language model. It extracts the code’s local semantic features as node embedding in the control flow graph by using sequence embedding. Secondly, CSGVD uses graph neural networks to extract the structured information of the graph. Finally, we propose a mean biaffine attention pooling, M-BFA, to better aggregate node information as a graph’s feature representation. The experimental results show that CSGVD outperforms the existing state-of-the-art models and obtains 64.46% accuracy on the real-world benchmark dataset from CodeXGLUE for vulnerability detection.

An Authentication Method for AMBTC Compressed Images Using Dual Embedding Strategies

In this paper, we proposed an efficient authentication method with dual embedment strategies for absolute moment block truncation coding (AMBTC) compressed images. Prior authentication works did not take the smoothness of blocks into account and only used single embedding strategies for embedment, thereby limiting image quality. In the proposed method, blocks were classified as either smooth or complex ones, and dual embedding strategies used for embedment. Respectively, bitmaps and quantized values were embedded with authentication codes, while recognizing that embedment in bitmaps of complex blocks resulted in higher distortion than in smooth blocks. Therefore, authentication codes were embedded into bitmaps of smooth blocks and quantized values of complex blocks. In addition, our method exploited to-be-protected contents to generate authentication codes, thereby providing satisfactory detection results. Experimental results showed that some special tampering, undetected by prior works, were detected by the proposed method and the averaged image quality was significantly improved by at least 1.39 dB.

DDS: A Delay-Based Differentiated Service Virtual Network Embedding Algorithm

Network virtualization (NV) is considered a promising technology that may solve the problem of Internet rigidity. The resource competition of multiple virtual networks for shared substrate network resources is a challenging problem in NV called virtual network embedding (VNE). Existing approaches do not consider the differences between multi-tenant requests and adopt a single embedding method, resulting in poor performance. This paper proposes a virtual network embedding algorithm that distinguishes the network types requested by tenants. This method divides virtual network requests into ordinary requests and delay-sensitive requests according to the delay constraints, provides personalized mapping strategies for different networks, and flexibly responds to the resource requirements and quality of service (QoS) requirements of the virtual network. The simulation results show that, compared with other algorithms, the proposed algorithm improves the request acceptance ratio by about 2% to 15% and the substrate network resources are more effectively utilized.

Deep multiview learning to identify imaging-driven subtypes in mild cognitive impairment

BackgroundIn Alzheimer’s Diseases (AD) research, multimodal imaging analysis can unveil complementary information from multiple imaging modalities and further our understanding of the disease. One application is to discover disease subtypes using unsupervised clustering. However, existing clustering methods are often applied to input features directly, and could suffer from the curse of dimensionality with high-dimensional multimodal data. The purpose of our study is to identify multimodal imaging-driven subtypes in Mild Cognitive Impairment (MCI) participants using a multiview learning framework based on Deep Generalized Canonical Correlation Analysis (DGCCA), to learn shared latent representation with low dimensions from 3 neuroimaging modalities.ResultsDGCCA applies non-linear transformation to input views using neural networks and is able to learn correlated embeddings with low dimensions that capture more variance than its linear counterpart, generalized CCA (GCCA). We designed experiments to compare DGCCA embeddings with single modality features and GCCA embeddings by generating 2 subtypes from each feature set using unsupervised clustering. In our validation studies, we found that amyloid PET imaging has the most discriminative features compared with structural MRI and FDG PET which DGCCA learns from but not GCCA. DGCCA subtypes show differential measures in 5 cognitive assessments, 6 brain volume measures, and conversion to AD patterns. In addition, DGCCA MCI subtypes confirmed AD genetic markers with strong signals that existing late MCI group did not identify.ConclusionOverall, DGCCA is able to learn effective low dimensional embeddings from multimodal data by learning non-linear projections. MCI subtypes generated from DGCCA embeddings are different from existing early and late MCI groups and show most similarity with those identified by amyloid PET features. In our validation studies, DGCCA subtypes show distinct patterns in cognitive measures, brain volumes, and are able to identify AD genetic markers. These findings indicate the promise of the imaging-driven subtypes and their power in revealing disease structures beyond early and late stage MCI.