In federated learning, model aggregation is performed at a common server leading to a single point of failure, which aggravates with increasing number of clients causing large network delays that may slow down the entire system. In such cases, hierarchical aggregation is performed with edge devices deployed at different geographical locations. Typically, edge devices employ online learning algorithms for processing the incoming data stream and so, one-way aggregation strategies might end up overwriting the local model updates after the clients receive the aggregated model. Hence, we propose a two-way aggregation strategy wherein models are aggregated both on the client and server. Also, since the incoming distribution may be temporal in nature, we need to dynamically update the aggregation tree. We quantify the non IIDness among different geo-spatially distributed dataspouts without compromising data privacy. We propose a novel dynamic re-clustering algorithm which is triggered by the intracluster and inter-cluster divergence. We implement the above algorithm in an end-to-end FL system that can perform two-step aggregation for aggregation of parent-side and child-side models in the aggregation tree to preserve the model updates in asynchronous execution of the system. Using benchmark datasets, we demonstrate that our proposed framework achieves 95% accuracy, which we plan to further improve.

The creation of personalized anatomical digital twins is important in the fields of medicine, computer graphics, sports science, and biomechanics. To observe a subject's anatomy, expensive medical devices (MRI or CT) are required and the creation of the digital model is often time-consuming and involves manual effort. Instead, we leverage the fact that the shape of the body surface is correlated with the internal anatomy; e.g. from surface observations alone, one can predict body composition and skeletal structure. In this work, we go further and learn to infer the 3D location of three important anatomic tissues: subcutaneous adipose tissue (fat), lean tissue (muscles and organs), and long bones. To learn to infer these tissues, we tackle several key challenges. We first create a dataset of human tissues by segmenting full-body MRI scans and registering the SMPL body mesh to the body surface. With this dataset, we train HIT (Human Implicit Tissues), an implicit function that, given a point inside a body, predicts its tissue class. HIT leverages the SMPL body model shape and pose parameters to canonicalize the medical data. Unlike SMPL, which is trained from upright 3D scans, MRI scans are acquired with subjects lying on a table, resulting in significant soft-tissue deformation. Consequently, HIT uses a learned volumetric deformation field that undoes these deformations. Since HIT is parameterized by SMPL, we can repose bodies or change the shape of subjects and the internal structures deform appropriately. We perform extensive experiments to validate HIT's ability to predict a plausible internal structure for novel subjects. The dataset and HIT model are available at https://hit.is.tue.mpg.de to foster future research in this direction.

Despite impressive performance for high-level downstream tasks, self-supervised pre-training methods have not yet fully delivered on dense geometric vision tasks such as stereo matching or optical flow. The application of self-supervised concepts, such as instance discrimination or masked image modeling, to geometric tasks is an active area of research. In this work, we build on the recent cross-view completion framework, a variation of masked image modeling that leverages a second view from the same scene which makes it well suited for binocular downstream tasks. The applicability of this concept has so far been limited in at least two ways: (a) by the difficulty of collecting real-world image pairs – in practice only synthetic data have been used – and (b) by the lack of generalization of vanilla transformers to dense downstream tasks for which relative position is more meaningful than absolute position. We explore three avenues of improvement. First, we introduce a method to collect suitable real-world image pairs at large scale. Second, we experiment with relative positional embeddings and show that they enable vision transformers to perform substantially better. Third, we scale up vision transformer based cross-completion architectures, which is made possible by the use of large amounts of data. With these improvements, we show for the first time that state-of-the-art results on stereo matching and optical flow can be reached without using any classical task-specific techniques like correlation volume, iterative estimation, image warping or multi-scale reasoning, thus paving the way towards universal vision models.

Masked Image Modeling (MIM) has recently been established as a potent pre-training paradigm. A pretext task is constructed by masking patches in an input image, and this masked content is then predicted by a neural network using visible patches as sole input. This pre-training leads to state-of-the-art performance when finetuned for high-level semantic tasks, e.g. image classification and object detection. In this paper we instead seek to learn representations that transfer well to a wide variety of 3D vision and lower-level geometric downstream tasks, such as depth prediction or optical flow estimation. Inspired by MIM, we propose an unsupervised representation learning task trained from pairs of images showing the same scene from different viewpoints. More precisely, we propose the pretext task of cross-view completion where the first input image is partially masked, and this masked content has to be reconstructed from the visible content and the second image. In single-view MIM, the masked content often cannot be inferred precisely from the visible portion only, so the model learns to act as a prior influenced by high-level semantics. In contrast, this ambiguity can be resolved with cross-view completion from the second unmasked image, on the condition that the model is able to understand the spatial relationship between the two images. Our experiments show that our pretext task leads to significantly improved performance for monocular 3D vision downstream tasks such as depth estimation. In addition, our model can be directly applied to binocular downstream tasks like optical flow or relative camera pose estimation, for which we obtain competitive results without bells and whistles, i.e., using a generic architecture without any task-specific design.

Abstract Rapid advancements in Information and Communication Technologies (ICT) and artificial intelligence (AI) applications permeating to all spheres of life, including medical prognosis, have led modern clinical systems to tread the path of advanced Internet of Medical Things (IoMT) by infusing advanced learning technologies, particularly deep learning. Automated diagnosis of malarial infection using AI‐enabled IoMT holds the promise of sustainable prognosis by reducing diagnosis error significantly with improved recognition accuracy. Existing automated diagnostic systems usually employ classical deep learning models wherein setting parameter values such as automatic learning rate selection, weight management etc. are a major concern. To address these issues, this paper proposes a collaborative ensemble AI‐enabled IoMT automated diagnosis model to classify malaria parasitized from microscopic images. The proposed model consists of two main stages. In the first stage, a Snapshot ensemble learning model is conjured upon by a combination of three distinct layers of Convolutional, Batch Normalization, and Relu networks; that alters the learning rate aggressively during training phase thus providing different network weights that gives multiple models by training a single model. In the second stage, an ensemble of three transfer learning models is constructed, and finally the average ensemble result is obtained. The learning rates at both these stages are empirically selected through Cosine Annealing. Experiment on the malaria parasite image dataset demonstrates the superiority of the proposed model with respect to a baseline algorithm.

CockroachDB is an open-source database, providing transactional access to data in a distributed setting. CockroachDB employs a multi-version timestamp ordering protocol to provide serializability. This provides a simple mechanism to enforce serializability, but the static timestamp allocation scheme can lead to a high number of aborts under contention. We aim to reduce the aborts for transactional workloads by integrating a dynamic timestamp ordering based concurrency control scheme in CockroachDB. Dynamic timestamp ordering scheme tries to reduce the number of aborts by allocating timestamps dynamically based on the conflicts of accessed data items. This gives a transaction higher chance to fit on a logically serializable timeline, especially in workloads with high contention.

Cloud-based data-intensive applications have to process high volumes of transactional and analytical requests on large-scale data. Businesses base their decisions on the results of analytical requests, creating a need for real-time analytical processing. We propose Janus, a hybrid scalable cloud datastore, which enables the efficient execution of diverse workloads by storing data in different representations. Janus manages big datasets in the context of datacenters, thus supporting scaling out by partitioning the data across multiple servers. This requires Janus to efficiently support distributed transactions. In order to support the different datacenter requirements, Janus also allows diverse partitioning strategies for the different representations. Janus proposes a novel data movement pipeline to continuously ensure up to date data between the different representations. Unlike existing multi-representation storage systems and Change Data Capture (CDC) pipelines, the data movement pipeline in Janus supports partitioning and handles both distributed transactions and diverse partitioning strategies. In this paper, we focus on supporting Online Transaction Processing (OLTP) and Online Analytical Processing (OLAP) workloads, and hence use row and column-oriented representations, which are the most efficient representations for these workloads. Our evaluations over Amazon AWS illustrate that Janus can provide real-time analytical results, in addition to processing high-throughput transactional workloads.

Variety of data has led to the fall of the "One size fits all" paradigm in databases. Applications now store data in multiple data representations, rather than just storing the data in a single relational database. However, applications only provide consistency semantics at the boundaries of a single datastore, corresponding to specific representation. This can lead to semantically inconsistent executions, which can have undesirable consequences for the application users. We propose, Typhon, a multi-representation data processing framework, which defines consistency semantics across multiple representations. Typhon introduces the notion of entities, which link the data present in diverse representations, and defines implicit causal guarantees, which capture the logical order intended by the application user. Typhon lays the foundation of a consistency model for multi-representation data, over which applications can build their protocols and reason about the consistency semantics of their existing solutions. We then propose a protocol, Cerberus, which provides implicit causal guarantees defined by Typhon's multi-representation consistency framework. Our evaluations show that Cerberus's performance is close to the solution providing no consistency guarantees across multiple representations, and outperforms baseline locking protocols providing Typhon's consistency guarantees across multiple data representations.

Internet of Things (IoT) applications like smart cars, smart cities and wearables are becoming widespread and are the future of the Internet. One of the major challenges for IoT applications is efficiently processing, storing and analyzing the continuous stream of incoming data from a large number of connected sensors. We propose a multi-representation based data processing architecture for IoT applications. The data is stored in multiple representations, like rows, columns, graphs which provides support for diverse application demands. A unifying update mechanism based on deterministic scheduling is used to update the data representations, which completely removes the need for data transfer pipelines like ETL (Extract, Transform and Load). The combination of multiple representations, and the deterministic update mechanism, provides the ability to support real-time analytics and caters to IoT applications by minimizing the latency of operations like computing pre-defined aggregates.

Web-based applications face unprecedented workloads demanding the processing of a large number of events reaching to the millions per second. That is why developers are increasingly relying on scalable cloud platforms to implement cloud applications. Chariots exposes a shared log to be used by cloud applications. The log is essential for many tasks like bookkeeping, recovery, and debugging. Logs offer linearizability and simple append and read operations of immutable records to facilitate building complex systems like stream processors and transaction managers. As a cloud platform, Chariots offers fault-tolerance, persistence, and high-availability, transparently. Current shared log infrastructures suffer from the bottleneck of serializing log records through a centralized server which limits the throughput to that of a single machine. We propose a novel distributed log store, called the Fractal Log Store (FLStore), that overcomes the bottleneck of a single-point of contention. FLStore maintains the log within the datacenter. We also propose Chariots, which provides multi-datacenter replication for shared logs. In it, FLStore is leveraged as the log store. Chariots maintains causal ordering of records in the log and has a scalable design that allows elastic expansion of resources.

For data-intensive applications with many concurrent users, modern distributed main memory database management systems (DBMS) provide the necessary scale-out support beyond what is possible with single-node systems. These DBMSs are optimized for the short-lived transactions that are common in on-line transaction processing (OLTP) workloads. One way that they achieve this is to partition the database into disjoint subsets and use a single-threaded transaction manager per partition that executes transactions one-at-a-time in serial order. This minimizes the overhead of concurrency control mechanisms, but requires careful partitioning to limit distributed transactions that span multiple partitions. Previous methods used off-line analysis to determine how to partition data, but the dynamic nature of these applications means that they are prone to hotspots. In these situations, the DBMS needs to reconfigure how data is partitioned in real-time to maintain performance objectives. Bringing the system off-line to reorganize the database is unacceptable for on-line applications.

Cross datacenter replication is increasingly being deployed to bring data closer to the user and to overcome datacenter outages. The extent of the influence of wide-area communication on serializable transactions is not yet clear. In this work, we derive a lower-bound on commit latency. The sum of the commit latency of any two datacenters is at least the Round-Trip Time (RTT) between them. We use the insights and lessons learned while deriving the lower-bound to develop a commit protocol, called Helios, that achieves low commit latencies. Helios actively exchanges transaction logs (history) between datacenters. The received logs are used to decide whether a transaction can commit or not. The earliest point in the received logs that is needed to commit a transaction is decided by Helios to ensure a low commit latency. As we show in the paper, Helios is theoretically able to achieve the lower-bound commit latency. Also, in a real-world deployment on five datacenters, Helios has a commit latency that is close to the optimal.

The past decade has witnessed an increasing adoption of cloud database technology, which provides better scalability, availability, and fault-tolerance via transparent partitioning and replication, and automatic load balancing and fail-over. However, only a small number of cloud databases provide strong consistency guarantees for distributed transactions, despite decades of research on distributed transaction processing, due to practical challenges that arise in the cloud setting, where failures are the norm, and human administration is minimal. For example, dealing with locks left by transactions initiated by failed machines, and determining a multi-programming level that avoids thrashing without under-utilizing available resources, are some of the challenges that arise when using lock-based transaction processing mechanisms in the cloud context. Even in the case of optimistic concurrency control, most proposals in the literature deal with distributed validation but still require the database to acquire locks during two-phase commit when installing updates of a single transaction on multiple machines. Very little theoretical work has been done to entirely eliminate the need for locking in distributed transactions, including locks acquired during two-phase commit. In this paper, we re-design optimistic concurrency control to eliminate any need for locking even for atomic commitment, while handling the practical issues in earlier theoretical work related to this problem. We conduct an extensive experimental study to evaluate our approach against lock-based methods under various setups and workloads, and demonstrate that our approach provides many practical advantages in the cloud context.