Contact current seminar convener:
Telephone (+61 8) 6488 3449
Our School runs a research seminar series to which all interested parties are invited to attend.
The usual seminar time is 11:00am on Friday mornings in Room 1.24.
In this talk, Li will briefly introduce the application background of keyword search over structured data. And then, he will present the technical challenges of dealing with keyword search over structured data and overview the summary of his research on keyword search in recent years. At last, Li will present his work - keywords-based social network data analysis, published in IEEE ICDE, 2014.
Dr. Jianxin Li is a Senior Lecturer in School of CSSE, UWA. His research area and interest are in XML/graph data management and data analysis, with a particular focus on the target-driven data processing and analytical technologies. He has published 40+ papers in top international conferences such as IEEE ICDE and EDBT, and premier journals such as IEEE TKDE, Information Systems and the Computer Journal.
UWA recently entered into an agreement with a tech startup called Fleet Engineering under the UWA Innovation Quarter's MO4U program. The agreement has led to the company co-locating in the UWA Computer Science building and creating exciting opportunities for collaborative research, student internships, mentoring student projects and teaching support. Fleet Engineering have a team of 4 full time and 3 part time staff, developing proprietary software for logistics and related industries to help manage assets, jobs and drivers to improve safety, productivity and customer experience. In this seminar we will tell the Fleet Engineering story and discuss the opportunities that arise from this agreement. The talk will be in 3 sections of approximately 15 minutes each, with 5 minutes to answer questions after each.
Samuel Hall - An application to identify the Source of Project Error
The key to a well-executed software project is the plan and adherence to it, a sign of a poor or immature software process is the inability of an organisation to meet the targets of a project or fulfil the customer's requirements. The goal for any software organisation should be to improve the planning, estimation and execution of the software process to ensure projects are well executed. Software Process Improvement (SPI) has been shown to be effective in ensuring the ongoing success of software development, however existing SPI methods can be costly or require specific knowledge.
A Rules Based Expert System using a set of equations for Partitioning the Mean Square Error (PAMESE) has been devised to allow project managers to quickly and easily identify issues within a software process using existing metrics of measurement. The application examines past project data and advises the project manager of causes of error within the process prioritising low effort high impact improvements as well as reporting the total error across the data set. Simulations of common issues within software organisations will be used to explore the effectiveness of this statistical aid to Software Process Improvement.
Areej Alsini - Comparing Three Convolutional Neural Network Architectures in Single object Image Classification using DIGITS
Deep learning algorithms have been widely applied in many computer vision applications especially in image classification. There is a dramatic improvement in the image classification performance. A common type of deep learning networks is the Convolutional Neural Networks often abbreviated as CNN or ConvNet. Various deep convolutional architectures, such as LeNet, AlexNet and GoogLeNet, have been proposed in the literature. Among these networks, LeNet, proposed in 1998, was one of the earliest deep learning architectures for classifying numerical digits (0-9). This five layers network is one of the fundamental models that is studied in this thesis. For object classification, the AlexNet with 8 layers and the GoogLeNet with 22 layers network were the state of the art networks with the minimum error rates in 2012 and 2014, respectively. There is a lack of systematic study to compare the three networks. The DIGITS platform was developed by NVIDIA in 2015. These models were implemented in the DIGITS version 3 framework which supports image classification and integrates both Torch 7 and Caffe with friendly user interface.
In this project, the three convolutional deep learning architectures, namely the LeNet, AlexNet and GoogLeNet, implemented under the DIGITS framework are studied and compared in terms of their architectures, training time, disk usage and classification performance. To speed up the training process, a GPU, NVIDIA GEFORCE GTX 960M, is used. The experiments were conducted on 2 standard benchmark datasets: Cifar-10 and Cifar-100. In all the experiments, the AlexNet was found to outperform the other two networks.
Jun Yang - 3D Object Recognition
This research is an attempt to address the issue of RGB-D image-based object recognition by using the SUN RGB-D dataset, which is a dataset containing 10,335 RGB-D images with dense annotations in both 2D and 3D, for both objects and rooms. The object recognition problem can be defined as a process of labelling object categories based on models of known objects. Formally, given an image which contains one or more objects and a set of labels corresponding to a set of models known to the system, the system is supposed to assign correct label or labels to regions in the image. In this research, a learned feature descriptor is proposed based on a convolutional k-means descriptor. The basic process to extract feature descriptor can be divide into three steps. First, the feature responses must be learned extracting the image structure of patches. Second, an interest point detector is run on the input image. Third, the convolutional feature descriptor responses are extracted around interest points. At last, a non-linear SVM classifier will be trained to classify the different types of feature responses into classes. There are several parameters that can be changed in this work: whether to use whitening or not, the number of features k, the receptive field size. The experiments with different parameters showed different results. And the use of depth information contributed significant information to the recognition.
Monte Masarei - Modelling the Targeted Removal of Micro-Cracks in Bone
Microscopic cracks form in bone due to the stresses and strains associated with normal daily activity. Bone remodelling is the body's solution to this problem, where damaged or underused bone is removed and replaced by healthy bone. The driving force behind this process is the bone multicellular unit (BMU) which is a long cylindrical group of cells consisting of osteoclasts at the BMU head (bone removal cells) and osteoblasts at the BMU tail (bone remodelling cells). The BMU tunnels through bone removing old bone at its head and replacing it with new bone at its tail, leaving behind it a cylindrical section of bone called an osteon.
It has been shown that bone remodelling occurs preferentially in the direction of principal stress. Recently studies have suggested BMU's can deviate from this path to target micro damage. Although a lot is known about individual biological and chemical processes thought to play a role in bone remodelling, the action of targeted bone remodelling is still not completely understood. This study integrates a mathematical model with bone biology to simulate target remodelling. A two dimensional cellular automata model was used to simulate diffusion and chemical interaction of molecules thought to be important in controlling targeted remodelling. BMU movement in a static chemical field was then modelled as the result of chemical gradients and concentration. Simulation managed to replicate BMU paths found in real human bone suggesting chemical signalling plays a key role in targeted remodelling.