computer science
OPENS DOORS
2016 Postgrad Symposium
Prizes following the symposium:The publication prize went to Feiyang Liu
The presentation prize went jointly to Lahiru Ariyasinghe and Joshua La Pine.
The symposium organisers were Aleksei Fedorov and Ali Knott.
DATE: Monday 17th October, 2016
VENUE: Room G34, Owheo Building, 133 Union St East, Dunedin
SCHEDULE:
Note: the overall presentation time is 20m: 15m talk and 5m for questions.
Abstracts
Wen Yang
Optical Network on Chips (ONoC) have become the main stream for Chip Multi-Processors(CMPs) design, due to the high-bandwidth capacity and high reliability with low propagation loss. Multicast, a one-to-many communication pattern, is widely used in barrier/clock synchronization, multithreading programs and cache coherence protocols for CMPs. Routing and wavelength assignment has a strong impact on several non-functional requirements of a NoC-based system. Performance, reliability, energy consumption, power dissipation, thermal aspects, and fault tolerance represent just a short list of the major common metrics affected by the routing and wavelength assignment. Even though several multicast routing algorithms have been proposed for ONoC, most of them consider only one multicast request. Few effective routing and wavelength assignment algorithms designed for multiple multicast requests. In my work, I will determine the condition on which given multicast assignments can be embedded in optical network on chip. Then, I will propose a novel multicast routing algorithm based on this condition to accommodate any given multicast requests using minimum wavelengths.
Feiyang Liu
The continuous developing of manufacturing technology turns many-core processors into reality. However, only a small number of cores can work at the same time due to the strict power budget. This phenomenon is called dark silicon. Optical Network on Chip is a chip-scale communication network which transmits date between different cores using light. Our research is to design a dark silicon aware routing and wavelength allocation scheme for Optical Network on Chip. The proposed scheme can provide non-blocking communication between any active cores by using the minimal number of wavelengths. Thus, the power consumption and hardware cost can be significantly decreased.
Leila Eskandary
Recently, the growth in the amount of data generated by new applications such as social networks, stock trading and monitoring has caused the emergence of new generation of Data Stream Management Systems (DSMSs). A DSMS is designed to process continuous and unending data as it arrives without storing it first. Task allocation policies in DSMSs have a significant impact on performance metrics such as data processing latency, maximal memory requirements for processing and system throughput. In this talk, I will discuss Apache Storm framework, outline existing scheduling algorithms designed for Storm and present a new scheduling algorithm that can efficiently distribute the tasks across a heterogeneous Storm cluster
Lahiru Ariyasinghe
HTTP Adaptive Streaming (HAS) is quickly becoming the de facto standard for Over-The-Top (OTT) video delivery. It enables the continuous delivery of live or Video-on-Demand (VoD) content without the need of guaranteed bandwidth. This relies on the client-side player being proposed a range of video quality levels from the server on a fragment basis. The client-side player then selects the next fragment according to the measured bandwidth with the server. However, it's clear that the existing HAS players are still at their infancy. When multiple players compete over a common bottleneck link, they often fail to determine their respective fair share of bandwidth. Ultimately, this leads to undesirable oscillations in the requested video quality and unfairness among the requested video qualities. In my talk, I will present the applicability of a coordinated bandwidth shaping technique (a control plane approach) that enables a set of HAS players to achieve a higher stability and fairness compared to the current state of the art.
Abbas Arghavani
Since radio links in wireless body area networks (WBAN) commonly experience highly time-varying attenuation due to topology instability, communication protocols with fixed transmission power cannot produce very good performance in terms of energy consumption and communication reliability. This talk presents a power-adaptive communication protocol based on channel prediction to save energy, improve reliability, and reduce interference. We first demonstrated that existing channel models either cannot accurately predict channel burstiness or have high prediction complexity, and then proposed a new model with high predication accuracy and low complexity. Such a model allows the dynamic power adjustment at a per-transmission level. Since the data rate is low in many applications of WBANs, energy can be saved by postponing the packet transmission if the channel is currently bad but will come back to good within the packet deadline. We developed an optimal scheduling policy to dynamically schedule packet transmission and select the best time and transmission power to transmit each packet. We evaluate our scheme through extensive trace-driven simulations, and results demonstrate that our scheme can self-learn the channel burstiness patterns, and choose the best transmission power to reduce energy consumption and improve communication reliability. With retransmissions, our scheme can achieve 100% communication reliability but consumes only a half of the energy in comparison with communication protocols with fixed transmission power.
Lewis Baker
Fast power line detection in images is useful in photogrammetry applications such as measuring wire tension and sag. To make these measurements, images of entire power line spans must be used which may include large amounts of curvature. Previous work in power line detection has focused on aerial or close proximity images where no power line curvature is visible. This talk assesses feasibility of ground-based imaging utilising smartphone cameras together with fast and robust power line detection using two common Hough transform techniques, and a line tracing algorithm.
Aleksei Fedorov
In the era of massive MIMO and very large array antennas with high-resolution, wireless channel models have to scrutinize the detailed space features of a surrounding environment. The existing models such as WINNER, 3GPP and IEEE 802.11 are not appropriate for validating and evaluating of new concepts and approaches for 4G/5G standards because they cannot provide a channel with realistic spatial characteristics. Meanwhile, spatial characteristics are the important data for new algorithms of MIMO antennas. A number of map-based channel models were proposed, which exploit the ray tracing approach and consider simplified 3D shapes of buildings and other objects to simulate the signals propagation effects such as diffraction, specular reflection, diffuse scattering, and etc. However, in the case of specular reflection, all the models utilize vertical and horizontal planes, whereas, a lot of reflecting surfaces are inclined in a real medium. Even a small displacement of a surface can significantly change the channel.
Kevin Xiao
The lossy channel links in the VANET make the design of a reliable broadcast scheme challenging. The broadcast requirements in the VANET, in terms of reliability, latency and network load, often conflict with each other. Existing solutions always try to balance the unpredictable delay in the duplication flooding and the poor reliability when averting the broadcast storm, rendering the problem unsolved. In this paper, we proposed a decentralized cooperative broadcast scheme with slot scheduling (DCBS), in which all vehicles jointly select and piggyback some received messages to help other vehicles recover the lost data. Slotted-CSMA is also employed in the cooperative piggybacking, and each vehicle explores the optimal piggybacking paths according to its local network information. The time slots are scheduled in terms of the optimal piggybacking paths to enhance the broadcast reliability. Simulation results show that the proposed scheme can achieve a higher performance in broadcast reliability, latency, and network load, comparing with other existing schemes.
Invited Speaker: Clinton Golding (HEDC)
A thesis is essentially a written work, and so to get a PhD or a Masters you must be a writer as well as a researcher. But writing can be difficult, stressful and slow. In this session we will discuss some tools and strategies so you can use to write more, write more easily and write better. We may do some practical writing exercises so please bring topics you want to write about and draft writing ready for refining and editing, which you are willing to share with others.
Nicolas Hananeia
We seek to model synaptic plasticity in the CA1 pyramidal cell, a principal neuron located in the hippocampus. As the hippocampus is responsible for recording, consolidation, and recall of episodic memory, the CA1 cell's prominent position in the hippocampus has made it an object of intense study. We follow from work during a previous Master's thesis (Hananeia & Benuskova, 2016) in which the Izhikevich model of a spiking neuron was used in conjunction with the Benuskova & Abraham (2007) model of synaptic plasticity to model the effects of high-frequency stimulation on the dentate gyrus granule cell. Here, we aim to use these same models modified for the CA1 pyramidal cell. When this model is complete, we aim to focus on reproducing the results in the synaptic plasticity study of Dong et al. (2008) with this model. This particular study examines heterosynaptic plasticity: that is, the modification of synaptic strength in a pathway adjacent to one undergoing stimulation. The study compares results obtained under different types of anaesthesia; we hope test in-silico the authors' hypothesis that urethane anaesthesia (as opposed to pentobarbital) maintains the natural input patterns to the cell yielding a different result. In preparation for this, we have reviewed recent literature on the neurobiology of the hippocampus and CA1 pyramidal cell, as well as on experimental and theoretical studies of synaptic plasticity.
Daniel Slack
Chris Gorman
We present an investigation of the potential use of Hopfield networks to learn neurally plausible, distributed representations of category prototypes. Hopfield networks are dynamical models of autoassociative memory which learn to recre- ate a set of input states from any given starting state. These networks, how- ever, will almost always learn states which were not presented during training, so called spurious states. Historically, spurious states have been an undesire- able side-effect of training a Hopfield network and there has been much research into detecting and discarding these unwanted states. However, we suggest that some of these states may represent useful information. It would be desirable for a memory system trained on multiple instance tokens of a category to extract a representation of the category prototype. We present an investigation showing that Hopfield networks are in fact capable of learning category prototypes as strong, stable, attractors without being explicitly trained on them. We also ex- pand on previous research into the detection of spurious states in order to show that it is possible to distinguish between trained, spurious, and prototypical attractors
Joshua La Pine
Timbre is the term used to describe the elements of sound that are neither pitch nor loudness. It refers to the core perceptual characteristics of a sound, the elements that allow one to distinguish between a guitar and a piano, for example. This non definition is due to our lack of understanding about what perceptually meaningful elements sound is comprised of. If we could comprehensively define what constitutes timbre then synthesis and classification of any sound would be possible. This problem is hugely difficult due to timbre being perceptual in nature.
Rassoul Mesbah
We propose a fully automatic method for segmenting myelin and axon from microscopy images of excised mouse spinal cord based on Convolutional Neural Networks (CNNs) and Deep Convolutional Encoder-Decoder. We compare a two-class CNN, multi-class CNN, and multi-class deep convolutional encoder- decoder with traditional methods. The CNN method gives a pixel-wise accuracy of 79.7% whereas an Active Contour method gives 59.4%. The encoder-decoder shows better performance with 82.3% and noticeably shorter classification time than CNN methods.
Tapabrata Chakraborti
Fine-grained Visual Categorization deals with object detection and recognition with subtle inter-class differences, a case in point being automated species recognition from images. This talk introduces the problem, discusses some popular approaches, suggests a possible alternative approach and presents some initial experimental results. The current work also introduces a new benchmark dataset NZBrids v1.0 (developed at CS dept, Otago Univ) containing 600 images from 30 endemic NZ bird species.
