星空传媒

Sedenion Valued Neural Networks

In this project, a sedenion-valued neural network (Mc-SVNN) and its learning algorithm are proposed and used to solve several multi-time-steps and multi-task forecasting problems. The Mc-SVNN contains two components: a sedenion-valued neural network that represents the cognitive component, and a metacognitive component, which serves to self-regulate the learning algorithm. At each epoch, the metacognitive component decides what, how, and when learning occurs. The algorithm deletes unnecessary samples and stores only those that are used. This decision is determined by the sedenion magnitude and the 15 sedenion phases. The Mc-SVNN is applied to four real-world forecasting problems: USD-to-euro currency exchange rate forecasting, the sunspot number time series, power demand forecasting, traffic forecasting, and daily temperature prediction in Abu Dhabi. Compared to existing methods, the Mc-SVNN demonstrates superior performance in time-series forecasting while using a smaller number of parameters

Forecasting of Gridded Geo-Spatial Processes

We developed systems for solving several space-time forecasting problems like weather forecasting and traffic forecasting. Proposed systems incorporate ideas from multi-scale segmentation networks like U-Nets and achieved advanced positions in reputable international research competitions. For example, we achieved the 4th position in the NeurIPS 2021 Traffic4cast challenge and the 4th position in the ACM CIKM 2021 Weather4cast competition. 

Semantic Labelling of Remote Sensing Images

Semantic segmentation necessitates approaches that learn high-level characteristics while dealing with enormous amounts of data. Convolutional neural networks (CNNs) can learn unique and adaptive features to achieve this aim. However, these networks cannot analyze an entire scene efficiently due to the large size and high spatial resolution of remote sensing images. Recently, deep transformers have proven their capability to record global interactions between different objects in the image. In this project, we propose a new segmentation model that combines convolutional neural networks with transformers and show that this mixture of local and global feature extraction techniques provides significant advantages in remote sensing segmentation. In addition, the proposed model includes two fusion layers designed to efficiently represent multi-modal inputs and output of the network. The input fusion layer extracts feature maps summarizing the relationship between image content and elevation maps (DSM). The output fusion layer uses a novel multi-task segmentation strategy where class labels are identified using class-specific feature extraction layers and loss functions. Finally, a fast-marching method is used to convert unidentified class labels to their closest known neighbors. Our results demonstrate that the proposed methodology improves segmentation accuracy compared to state-of-the-art techniques.

Explainable early detection of Alzheimer’s disease using ROIs and an ensemble of 138 3D vision transformers

We propose a new approach to Alzheimer disease (AD) prediction and interpretable AI for medical images is proposed. Brain MRIs were segmented into 138 regions. An ensemble of 138 3D vision transformers was used next to predict Alzheimer's disease progression, with each vision transformer dedicated to processing one specific region. This approach achieves SOTA performance and easily identifies relevant regions for each of the classification problems (AD vs. CN, AD vs. MCI, MCI vs CN)

Deep Hypercomplex Networks for Spatiotemporal and in Orbit Data Processing

This project explores the innovative application of hypercomplex numbers (such as quaternions, octonions, and sedenions) in creating parameter-efficient neural networks. By leveraging these numbers, the study demonstrates a significant reduction in network parameters, up to 16 times in the case of sedenions, while enhancing performance, especially in processing spatiotemporal data.  Key deep learning components, including concatenation, activation functions, convolution, and batch normalization, are adapted to the hypercomplex domain.  The integration of a ResNet backbone with hypercomplex convolution in a U-Net configuration is successfully applied to weather and traffic forecasting problems, highlighting the potential of hypercomplex networks to outperform traditional real-valued models under a fixed parameter budget.