Visualization and visual computing use computer-supported, interactive, visual representations of (abstract) data to amplify cognition. In recent years data complexity concerning volume, veracity, velocity, and variety has increased considerably. This is due to new data sources as well as the availability of uncertainty, error, and tolerance information.

I’ll cover thee different RPC policies implemented on top of R2P2. Specifically, we’ll see how R2P2 enables efficient in-network RPC load balancing based on a novel  join-bounded-shortest-queue (JBSQ) policy. JBSQ lowers tail latency by centralizing pending RPCs in the middle box and ensures that requests are only routed to servers with a bounded number of outstanding requests. Then, I’ll talk about SVEN, an SLO-aware RPC admission control mechanism implemented as an R2P2 policy on P4 programmable switches. Finally, I’ll describe HovercRaft, a new approach to building fault-tolerant generic RPC services by integrating state-machine replication in the transport layer.

Network communication is a major bottleneck in large-scale distributed deep learning. To minimize the problem, many compressed communication schemes, in the form of quantization or sparsification, have been proposed. We investigate them from the Computer Systems perspective, under real-life deployments. We identify discrepancies between the theoretical proposals and the actual implementations, and analyze the impact on convergence.

This talk will introduce two novel models we developed for automatic HG generation in two different settings, positive-negative learning and positive-unlabeled learning. We demonstrate the efficacy of the proposed model on three real-world datasets constructed from biomedical publications.

In this dissertation, we consider extensions of dynamic programming for combinatorial optimization. We introduce two exact multi-objective optimization algorithms: the multi-stage optimization algorithm that optimizes the problem relative to the ordered sequence of objectives (lexicographic optimization) and the bi-criteria optimization algorithm that simultaneously optimizes the problem relative to two objectives (Pareto optimization).

The talk will consist of two parts. First an introduction to important subsurface structures in relation to hydrocarbon extraction are given, followed by an overview of techniques to segment out and visualize these structures. The talk will include topics such as ground truth visualization of measured seismic data, automated object extraction for getting computer assistance in segmenting important structures in the seismic data such as horizons and faults and creation, and visualizing and visual fusion of multiattribute seismic data using the GPU.

We developed extensions of dynamic programming which allow us (i) to describe the set of objects under consideration, (ii) to perform a multi-stage optimization of objects relative to different criteria, (iii) to count the number of optimal objects, (iv) to find the set of Pareto optimal points for the bi-criteria optimization problem, and (v) to study the relationships between two criteria.

The term P4 medicine has been coined almost a decade ago to indicate novel ways of early detection and prevention of diseases. P4 stands for personalized, predictive, preventive, and participatory, to indicate that a diagnosis or treatment is tailored to each individual patient, risk factors are identified early and addressed before manifestation, and individuals are actively involved into all processes. Often, P4 approaches are accompanied by the acquisition of large and complex medical imaging data, and demanding computational processes–especially, when it comes to cancer radiotherapy, which is a data heavy and visual computing rich process.

A traditional goal of algorithmic optimality, squeezing out operations, has been superseded because of evolution in architecture. Algorithms must now squeeze memory, data transfers, and synchronizations, while extra operations on locally cached data cost relatively little time or energy. Hierarchically low-rank matrices realize a rarely achieved combination of optimal storage complexity and high-computational intensity in approximating a wide class of formally dense operators that arise in exascale applications. They may be regarded as algebraic generalizations of the fast multipole method.

Research in visualization and computer graphics has developed techniques to geometrically model objects from our everyday life, or from various branches of industry and science, including modeling life-forms that are of submicron in size. These are not visible to the naked eye and most of us are unfamiliar with structures that form life organized in an assembly of biomolecules. Here visualization techniques can be of tremendous help to guide the viewers to familiarize themselves with what they see and make the engaging visual exploration of these complex structures to an intellectual enrichment.

Research and experimentation in various scientific fields are based on the knowledge and ideas from scholarly literature. The advancement of research and development has, thus, strengthened the importance of literary analysis and understanding. However, in recent years, researchers have been facing massive scholarly documents published at an exponentially increasing rate. Analyzing this vast number of publications is far beyond the capability of individual researchers. This dissertation is motivated by the need for large scale analyses of the exploding number of scholarly literature for scientific knowledge discovery and information retrieval. In the first part of this dissertation, the interdependencies between scholarly literature are studied. First, I develop Delve -- a data-driven search engine supported by our designed semi-supervised edge classification method. This system enables users to search and analyze the relationship between datasets and scholarly literature. Based on the Delve system, I propose to study information extraction as a node classification problem in attributed networks. Specifically, if we can learn the research topics of documents (nodes in a network), we can aggregate documents by topics and retrieve information specific to each topic (e.g., top-k popular datasets).

The talk will be pre-recorded and a youtube link will be provided to seminar participants. During the seminar time, Prof. Wonka will mainly answer some questions from the audience.

Biological knowledge is widely represented in the form of ontologies and ontology-based annotations. The structure and information contained in ontologies and their annotations make them valuable for use in machine learning, data analysis and knowledge extraction tasks. In this thesis, we propose the first approaches that can exploit all of the information encoded in ontologies, both formal and informal, to learn feature embeddings of biological concepts and biological entities based on their annotations to ontologies by applying transfer learning on the literature. To optimize learning that combines ontologies and natural language data such as the literature, we also propose a new approach that uses self-normalization with a deep Siamese neural network to improve learning from both the formal knowledge within ontologies and textual data. We validate the proposed algorithms by applying them to generate feature representations of proteins, and of genes and diseases.

Out-of-Core simulation systems often produce a massive amount of data that cannot fit on the aggregate fast memory of the compute nodes, and they also require to read back these data for computation. As a result, I/O data movement can be a bottleneck in large-scale simulations. Advances in memory architecture have made it feasible to integrate hierarchical storage media on large-scale systems, starting from the traditional Parallel File Systems to intermediate fast disk technologies (e.g., node-local and remote-shared NVMe and SSD-based Burst Buffers) and up to CPU’s main memory and GPU’s High Bandwidth Memory. However, while adding additional and faster storage media increases I/O bandwidth, it pressures the CPU, as it becomes responsible for managing and moving data between these layers of storage. Simulation systems are thus vulnerable to being blocked by I/O operations. The Multilayer Buffer System (MLBS) proposed in this research demonstrates a general method for overlapping I/O with computation that helps to ameliorate the strain on the processors through asynchronous access. The main idea consists in decoupling I/O operations from computational phases using dedicated hardware resources to perform expensive context switches. By continually prefetching up and down across all hardware layers of the memory/storage subsystems, MLBS transforms the original I/O-bound behavior of evaluated applications and shifts it closer to a memory-bound or compute-bound regime.

In my talk, I will present techniques that allow biologists to model a mesocale entity in a rapid way in the timeframe of a few minutes to hours. This way we have created the first complete atomistic model of the SARS-CoV-2 virion that we are these days sharing with the worldwide scientific community. Mesoscale represents a scalar gap that is currently not possible to accurately image with neither microscopy nor X-ray crystallography approaches. For this purpose, scientists characterize it by observations from the surrounding nanoscale and the microscale. From this information, it is possible to reconstruct a three-dimensional model of a biological entity with a full atomistic model. The problem is that these models are enormously large and are not possible to model with traditional methods from computer graphics within a reasonable time.

In this talk, we will discuss a new way of computing with quad meshes. It is based on the checkerboard pattern of parallelograms one obtains by subdividing a quad mesh at its edge midpoints. The new approach is easy to understand and implement. It simplifies the transfer from the familiar theory of smooth surfaces to the discrete setting of quad meshes. This is illustrated with applications to constrained editing of 3D models, mesh design for architecture and digital modeling of shapes which can be fabricated by bending flat pieces of inextensible sheet material.

In the lecture, we present a three-dimensional model for the simulation of signal processing in neurons. Part of this approach is a method to reconstruct the geometric structure of neurons from data measured by 2-photon microscopy. Being able to reconstruct neural geometries and network connectivities from measured data is the basis of understanding coding of motoric perceptions and long term plasticity which is one of the main topics of neuroscience. Other issues are compartment models and upscaling.

Dynamic programming is an efficient technique to solve optimization problems. It is based on decomposing the initial problem into simpler ones and solving these sub-problems beginning from the simplest ones. A conventional dynamic programming algorithm returns an optimal object from a given set of objects. We developed extensions of dynamic programming which allow us (i) to describe the set of objects under consideration, (ii) to perform a multi-stage optimization of objects relative to different criteria, (iii) to count the number of optimal objects,(iv) to find the set of Pareto optimal points for the bi-criteria optimization problem, and (v) to study the relationships between two criteria.

We developed and expanded novel methods for representation learning, predicting protein functions and their loss of function phenotypes. We use deep neural network algorithm and combine them with symbolic inference into neural-symbolic algorithms. Our work significantly improves previously developed methods for predicting protein functions through methodological advances in machine learning, incorporation of broader data types that may be predictive of functions, and improved systems for neural-symbolic integration. The methods we developed are generic and can be applied to other domains in which similar types of structured and unstructured information exist. In future, our methods can be applied to prediction of protein function for metagenomic samples in order to evaluate the potential for discovery of novel proteins of industrial value.  Also our methods can be applied to the prediction of loss of function phenotypes in human genetics and incorporate the results in a variant prioritization tool that can be applied to diagnose patients with Mendelian disorders.

In this seminar, I will present some of the work I have done on Continual Deep Learning, among the research topics at the Vision-CAIR group. Continual learning aims to learn new tasks without forgetting previously learned ones. This is especially challenging when one cannot access data from previous tasks and when the model has a fixed capacity as adopted in modern deep learning techniques.  Decreasing the gap towards human-level continual learning, we extended continual deep learning from multiple perspectives. The Hebb's learning theory from biology can be famously summarized as “Cells that fire together wire together.". Inspired by this theory from biology, we proposed Memory Aware Synapses (ECCV18) to quantify and reduce machine forgetting in a way that enables leveraging unlabeled data, which was not possible in former techniques. We later developed a Bayesian approach appearing at ICLR2020, where we explicitly modeled uncertainty parameters to orchestrates forgetting in continual learning. We showed in our ICLR2019 and ACCV18 works that task descriptors/ language can operate in continual learning visual tasks to improve learning efficiency and enable zero-shot task transfer. Beyond computer vision tasks, we recently developed an approach appearing at ICLR2020 we call "Compositional Language Continual Learning". We showed that disentangling syntax from semantics enables better compositional Seq2Seq learning and can significantly alleviate forgetting of tasks like machine translation.  In the talk, I will go over these techniques and shed some light on future research possibilities.

This talk discusses a body of work on exploiting the DSP Blocks in modern FPGAs to construct high-performance datapaths including the concept of FPGA overlays. It outlines work that established FPGAs as a viable virtualized cloud acceleration platform, and how the industry has adopted this model. Finally, it discusses recent work on incorporating accelerated processing in network controllers and the emerging concept of in-network computing with FPGAs. These strands of work come together to demonstrate the value of thinking about computing beyond the CPU-centric view that still dominates.

In this talk, I will present a line of work done at the Image and Video Understanding Lab (IVUL), which focuses on developing deep graph convolutional networks (DeepGCNs). A GCN is a deep learning network that processes generic graph inputs, thus extending the impact of deep learning to irregular grid data including 3D point clouds and meshes, social graphs, protein interaction graphs, etc. By adapting architectural operations from the CNN realm and reformulating them for graphs, we were the first to show that GCNs can go as deep as CNNs. Developing such a high capacity deep learning platform for generic graphs opens up many opportunities for exciting research, which spans applications in the field of computer vision and beyond, architecture design, and theory. In this talk, I will showcase some of the GCN research done at IVUL and highlight some interesting research questions for future work.

In this talk, I will use the Empirical Risk Minimization (ERM) problem as an example and show how to overcome these challenges. Particularly, I will first talk about how to overcome the high dimensionality challenge from the data for Sparse Linear Regression in the local DP (LDP) model. Then, I will discuss the challenge from the non-interactive LDP model and show a series of results to reduce the exponential sample complexity of ERM. Next, I will present techniques on achieving DP for ERM with non-convex loss functions. Finally, I will discuss some future research along these directions.

This talk is about building compilers for high-performance code generation. It has three parts. The first part is about Tiramisu (http://tiramisu-compiler.org/), a polyhedral compiler designed for generating highly efficient code for multicores and GPUs. It is the first polyhedral compiler that can match the performance of highly hand-optimized industrial libraries such as Intel MKL and cuDNN. The second part is about applying Tiramisu to accelerate deep learning (DNN) inference. In comparison to other DNN compilers, Tiramisu has two unique features: (1) it supports sparse DNNs; and (2) it can express and optimize general RNNs (Recurrent Neural Networks). The third part will present recent work on the problem of automatic code optimization. In particular, it will focus on using deep learning to build a cost model to explore the search space of code optimizations.