| Proj-1 |
Network reliability - Dr. Louis Petingi. A communication network modeled as a undirected (directed) graph G = (V, E) where V is the set of vertices and E is the set of edges of G. Given a set of terminal vertices K of V (also called participating nodes), where each edge e of G is assigned a probability of operation, the classical reliability measure, RK(G), introduced in the 1970s, is the probability that each pair of terminal vertices u and v, is connected by a path. The Diameter-constrained reliability RK(G,d) (DCR), a generalization of the classical reliability, and introduced in 2004, gives the probability that given a diameter bound d, for each pair of terminal vertices u and v, at least an operational path of length d or less that connects u and v. This is applicable when assessing, for example, the reliability for any packet-oriented network where links may fail and a "time-to-live" (TTL) limit is specified in the number of hops that can be traversed by any given packet (for instance IPv6 packets include a hop limit field). As the classical reliability measure does not capture the distances between the network’s vertices, the DCR can be applied to assess the probability of establishing or maintaining a connection by setting, for example, the diameter bound d equal to the maximum number of hops to be visited by a packet or request. As the evaluation of the DCR is an NP-hard problem, students implement exact evaluation algorithms to calculate the DCR, such as Moskowitz's Factoring Theorem, and parallelize it using MPI libraries in C++, to improve the computational complexity. Then reliability-increasing graph transformations such as Kelmans' topological surgery are explored within the context of the DCR to modify a given network into a more reliable one; this transformation was originally introduced within the frame of the classical reliability but never studied when diameter constraints are under consideration. |
| Proj-2 |
Parallel collision search for cryptographic hash functions - Dr. Xiaowen Zhang. A cryptographic hash function (short for hash function) takes a much longer input message of arbitrary length and outputs a very shorter fixed-length bit-string, called hash. Since a large domain is mapped to a smaller range, collisions (pairs of inputs are mapped to the same output) are inevitable. However, as required for a hash function, it should be computationally infeasible to find any two distinct inputs that hash to the same value, i.e., collision resistant. Hash functions are commonly used for data integrity in conjunction with digital signature. Parallel collision search for hash function is to find hash collisions in an efficient and effective way. |
| Proj-3 |
Performance Analysis of Parallel Particle Filters - Dr. Feng Gu. Particle filters are widely used in data assimilation of simulation systems, which are sample-based methods using Bayesian inference and sampling techniques to recursively predict the state of the system with given observations. Practically, a large number of particles are needed to guarantee the convergence of particles, especially for large-scale dynamic systems. Therefore, parallel particle filters are needed to improve the performance by deploying particles on different processing units. In the three main steps of particle filters, sampling and weight computation have no data dependency between processing units. However, resampling needs the global processing, resulting in communication costs. To decrease the communication cost, various particle routing policies are proposed, such as minimal transfer particle routing and maximal balance particle routing policy. In this work, we will systematically analyze the performance of parallel particle filters with various routing policies and show their effectiveness. |
| Proj-4 |
Automatic cell counting in microscopy images using convolutional neural network - Dr. Shuqun Zhang. Cell counting is an important process in the field of biological and pathological image analysis. It is still performed manually in many labs, and it is necessary to develop an automatic counting method to obtain accurate and consistent results. Most of the existing methods are based on object detection and segmentation, which cannot achieve satisfactory results due to low quality of images, noise, overlapping cells, and complex background and shapes in microscopy images. Modern machine learning methods especially deep learning have recently revolutionized the field of image processing and computer vision. This project will explore the use of deep learning approach in cell counting to obtain more robust results. We will evaluate the performance of deep convolutional neural networks on cell counting task, and compare the speed of GPU and CPU implementations. |
