Nanyang Technological University

School of Computer Science and Engineering (SCSE)

​​​​​Name of NTU Supervisor

Research Title

Assoc Prof​ Suresh Sundaram Multi-Scale Search Algorithms For Black-Box / Derivative-Free Optimization 
SSundaram@ntu.edu.sg ​This project aims to study multi-scale search optimization algorithms and benchmark its efficacy on real-world problems. To this end, we seek to consolidate several state-of-the-art multi-scale search algorithms in a single programming platform/interface, which consequently facilitates their use and comparison. The second goal of this study is to inject established evolutionary and large-scale strategies such as adaptive encoding and random embedding within the framework of multi-scale search methods.

Skills required: Matlab, Python (IPython, Matplotlib, …)

Outcomes: research papers are expected upon the successful completion of the project.

Assoc Prof​ Suresh Sundaram ​Efficient Computational Cognitive System For Collaborative Learning 
SSundaram@ntu.edu.sg This project aims to investigate efficient and scalable algorithms for collaborative learning, which are used in analyzing big data problems. The existing meta-cognitive learning algorithm has been tested in learning stream-of-data, but they have control parametes to tune. To this end, we seek to review collaborative algorithms and propose an efficient scalable algorithm for the same. The second goal of this internship is to incorporate such an algorithm into forecasting of wind.

Skills required: Matlab, Python (IPython, Matplotlib, …), C/C++

Outcomes: research papers are expected upon the successful completion of the project.

Assoc Prof​ Suresh Sundaram Predictive Analytics For Energy Systems
SSundaram@ntu.edu.sg This project aims to investigate efficient data-driven policies for energy systems.  Based on a huge dataset of historical records of energy systems, the project seeks to develop a distributed computing engine to build predictive model of critical quantites and measures in the system of interest. The second goal of this project is to employ these predictive models within the black-box optimization framework to derive efficient energy-saving policies.

Skills required: Matlab, Python (IPython, Matplotlib, …), Spark, SQL, Hadoop

Outcomes: research papers are expected upon the successful completion of the project.

Assoc Prof Kwoh Chee Keong​ Software tool for rendering DNA nano-structures
​The research on the structural stability of 3 Dimensional Origami is attracting many to consider the possibility to make useful tools for simulating thermodynamically stable DNA structures. Most of the major architectural monuments around the world are a conglomeration of basic geometric structures such as cylinders, spheres, tetrahedrons, cuboids. For example, the Big Ben Tower of London, is roughly a collection of four cuboids, one truncated pyramid and one pyramid in a fixed structural ratio. Combining these basic structures to form such complex structures is still a difficult task in 3 Dimensional Origami.


The ultimate key is to create such architectural monuments or other such structures using Origami. The primal aim is to formulate a software that can render such basic geometric structures of any input length from the existing data available about them. The software must learn from the statistical data available about a particular structure and approximate the result from it. The output will be the possible structure and its corresponding data such as the total base pairs required, total crossovers, edge/ radius length, error in the length, base pairs/turn etc.

After a successful rendering algorithm for these basic structures is deduced, the subsequent step is to input complex 3 Dimensional Structures such as architectural buildings. This software should be able to take into account of the structural ratio (in terms of size) of these basic structures and output the required stable complex structure by customising the sizes of the basic structures.

Prof​ Dusit Niyato Deep Reinforcement Learning for Resource Allocation in 5G Networks
​Reinforcement learning (RL) is used extensively in artificial intelligence domain to facilitate model free learning of the underlying environment. Deep learning on the other hand is used extensively by machine learning community for performing supervised or unsupervised learning via artificial neural networks. Deep reinforcement learning (DRL) is an enhanced version of traditional RL that uses deep learning to control practical systems. This project aims to propose efficient resource allocation algorithms based on DRL for 5G enabled wireless networks.
Prof. Thambipillai SrikanthanStereo Region-of-Interest Generation for Pedestrian Protection
astsrikan@ntu.edu.sg Advanced driver assistance systems (ADAS), and particularly pedestrian protection systems (PPS), have become an active research area due to the increasing need for improving traffic safety. Robust and real-time pedestrian detection is a challenging task as there is a need to take into account cluttered background, non-rigid appearance of pedestrians, etc. Region of Interest (ROI) generation is an essential component in vision based PPS. In this project, the candidate will evaluate stereo vision based ROI generation techniques for robust and real-time pedestrian detection in ADAS.

​Assoc Prof Anwitaman DattaImplementing and experimenting with novel erasure codes for distributed storage systems
anwitaman@ntu.edu.sg​This project will involve understanding and implementing several families of recently proposed erasure codes, followed by thorough experimentations in computer clusters to benchmark these codes under varied workloads. The student needs to have good understanding of number theory, particularly finite fields (in order to understand the code structures), and good programming skills in general, as well as with ​socket/network programming. Knowledge of distributed systems is desirable, but not necessary.
Assoc Prof Anwitaman DattaAdapting proof of data possession techniques for efficient RAID monitoring​
anwitaman@ntu.edu.sg​Proof of data possession (PDP) techniques are usually used to detect data integrity when data is stored/outsourced with an untrusted server (e.g., on the cloud). It is a probabilistic technique and thus highly efficient, and scales well with volume of data. The purpose of the project will be to adapt the ideas from PDP in the context of RAID (redundant array of independent disks) systems. The student should have very good linux kernel + C programming skills. Knowledge of RAID systems will be an additional advantage to have.​
​Assoc Prof Wen YonggangSocial TV Analytic System: Distributed Crawler
YGWen@ntu.edu.sgIn this project, we aim to develop a big-data platform for social TV analytics. The platform consists of multiple components, including a distributed data crawler, NoSQL, various data mining algorithms, and data visualization. The candidate will work with a team of researchers and engineers to develop this system. The part of the project is to survey existing technologies and develop a distributed data crawling module.
​Assoc Prof Wen Yonggang​Social TV Analytic System: Data Visualization
YGWen@ntu.edu.sg​In this project, we aim to develop a big-data platform for social TV analytics. The platform consists of multiple components, including a distributed data crawler, NoSQL, various data mining algorithms, and data visualization. The candidate will work with a team of researchers and engineers to develop this system. The part of the project is to survey existing technologies and develop a data visualization module.
Asst/P Anupam ChattopadhyayEvolvable Video COMPression(EVoComp) system design for online video compression
anupam@ntu.edu.sg   ​With increase in the resolution of videos, higher bandwidth is required for transmission, thereby necessitating the need for compression. The project aims at developement of an FPGA based system that is – adaptive, so that it adapts according to changes in video characteristics and – on-line so that it can compress/decompress a video stream as it is being transmitted/received. The EvoComp system will allow evolutionary algorithms to change the hardware configuration  in real time to change compression methodolgy depending on video characteristics. We aim to benchmark the developed system with the existing video compression technqiues.
Asst/P Anupam Chattopadhyay​Collaborative  energy-aware downloading for locally connected Android Mobile Devices
anupam@ntu.edu.sg​Energy conservation is a critical aspect of mobile devices. Services like downloading content are energy hungry. The project aims at developemnt of energy efficient techniques to manage downloading large files by collaborative downloading across locally connected (over Bluetooth or Wifi) Android devices to maximize throughput and at the same time taking into account the characteristics of the peers such as network usage, current cpu utilization, battery level, etc for partioning the download.
​Asst/P Erik Cambria​​Concept-Level Sentiment Analysis with SenticNet
cambria@ntu.edu.sg

This project is in the context of sentiment analysis or opinion mining. The opportunity to capture the opinions of the general public about social events, political movements, company strategies, marketing campaigns, and product preferences has raised growing interest both within the scientific community, leading to many exciting open challenges, as well as in the business world, due to the remarkable benefits to be had from marketing and financial prediction. 

 

The candidate will focus on improving the current version of the Sentic API (http://sentic.net/api), a publicly available resource for concept-level sentiment analysis. Possible improvements are listed below.

1. KNOWLEDGE EXPANSION

Enriching SenticNet with new concepts.

Possible ways to do this are:

a) write some code (preferably in python but java also can) to import concepts from other resources, e.g., Bing Liu's Opinion Lexicon, MPQA Subjectivity Lexicon, SentiWordNet, Harvard General Inquirer, LIWC, etc.

b) crowdsourcing, e.g., through surveys, quizzes or games

c) an ensemble of the above

2. CONSISTENCY CHECK

Finding errors in SenticNet, e.g., in terms of polarity or semantics. 

Possible ways to do this are:

a) do it manually

b) use the same code/resources mentioned before to find clashes (e.g., positive polarity in SenticNet but negative in SentiWordNet)

c) an ensemble of the above

3. FEATURE EXTENSION

Enriching SenticNet w/ new add-ons.

Possible ways to do this are:

a) add POS tags

b) add mood tags

c) add category tags

Possible extra activities for the candidate include working on any of the modules of our model for concept-level sentiment analysis (http://sentic.net/clsa-model.pdf).​

​Asst Prof Lam Siew KeiReliability-Aware High-Level Synthesis for Embedded Computing Platforms
​​siewkei_lam@pmail.ntu.edu.sg ​Single Event Upsets (SEU) induced faults poses a serious reliability problem in safety-critical embedded systems. Existing SEU mitigation techniques often incur unacceptable performance-power-cost overhead. The main objective of this project is to devise techniques that automatically synthesize reliable embedded systems from high-level design descriptions guided by constraints and target hardware architecture specifics. The candidate will investigate lightweight SEU mitigation techniques for fault tolerant custom hardware and realize the proposed techniques on state-of-the-art FPGA platforms. The candidate is expected to be experience in digital design and Verilog HDL, good foundation in data structures and algorithms, and good programming skills (C/C++).​
Asst Prof Lam Siew Kei​​​Accelerating Feature Detectors for Real-time Vision-based Applications
​​siewkei_lam@pmail.ntu.edu.sg ​​Feature detection is a fundamental step in many real time applications such as video tracking, visual SLAM and robotic navigation. However, existing implementations for feature detection is highly compute intensive and becomes a bottleneck for real time vision tasks. This project aims to develop hardware-efficient feature detectors on FPGA.​

​​Dr Ta Nguyen Binh DuongSecurity and privacy issues in IaaS clouds
donta@ntu.edu.sg The use of virtualization and resource multiplexing enable commercial cloud providers (e.g., Amazon EC2) to maximize utilization but at the same time introduce new security vulnerabilities. It has been demonstrated that malicious users could launch virtual machines (VM) which are placed co-resident (on the same physical machine) with the target VM. Such placement in turn may lead to cross-VM side-channel attacks to extract sensitive information from the target VM. In this project, we investigate solutions including but not limited to VM placement algorithms to ensure security and privacy for IaaS cloud data centers. This project assumes a good background in algorithms and programming skills (e.g., Java and Python).
Asst Prof Arijit KhanTowards Querying and Mining of Big-Graphs
arijit.khan@ntu.edu.sg

With the advent of the Internet, sources of data have increased dramatically, including the World-Wide Web, social networks, genome databases, knowledge graphs, medical and government records. Such data are often represented as graphs, where nodes are labelled entities and edges represent relations among these entities. Knowledge is hidden in the complex structure and attributes inside these networks. While querying and mining these linked datasets are essential for various applications, traditional graph algorithms may not be able to capture the rich semantics in these networks. In this project, we shall design novel techniques and systems for emerging graph workloads including graph pattern matching, approximate subgraph mining, similarity search, ranking and expert finding, aggregation and OLAP, uncertainty, and streaming.

 

*Project Duration: Minimum project duration is 6 months. Hence, this is suitable only for final year Bachelors or dual degree (integrated Bachelors and Masters) students for their final-year research project. The aim is to publish a paper in a top-tier database or data mining conference such as SIGMOD, KDD, VLDB, and ICDE.

*Please do not apply for this project if unable to commit for a period of  6 months

​Assoc Prof Chng Eng Siong​​Information extraction from Text
aseschng@ntu.edu.sg

In this work, we wish to examine the extraction of information from text.
E.g, its topic, its keywords, word cloud, its relationship to other paragraphs within a corpus.
An example of what can be achieved is seen in here:

Thomson Reuters Open Calais (*)

http://www.opencalais.com/

Detect entities, topic codes, events, relations and socialTags.

IBM Entity Extraction API (*)

http://www.alchemyapi.com/api

Detect entities of people, places, companies, topics, facts, relationships, authors, and languages.

As well as here: https://www.quora.com/What-are-the-available-APIs-for-NLP-Natural-Language-Processing

A potential project may be to link text transcription of a MOCC course such as DSP by
http://ocw.mit.edu/resources/res-6-008-digital-signal-processing-spring-2011/video-lectures/
to his book or other DSP books, allowing cross-referencing of text that is discussing similar topics.

​​​Assoc Prof Chng Eng SiongMachine Learning by TensorFlow https://www.tensorflow.org/
aseschng@ntu.edu.sg​Google has released its tensor flow codebase for machine learning
https://www.tensorflow.org/

This represents the state of the art system for machine learning. In this work we wish to examine its
use for the following applications
(choose one)

   a) signal enhancement (cleaning up noise e.g, https://www.youtube.com/watch?v=qvgsfP1J3Ss)
   b) classification of audio signals (similar in problem to classifying digits MNIST)
        https://www.tensorflow.org/versions/r0.9/tutorials/index.html
   c) classification of topic from a given text (topic classification)
   http://www.nltk.org/book/ch06.html

​Dr Ta Nguyen Binh DuongDistributed machine learning on public clouds
donta@ntu.edu.sgFor cloud gaming, fast provisioning of game servers is important to provide a good experience to players. To achieve this, usually the cloud gaming provider will need to prepare virtual machine images containing the desired games in advance. On the other hand, installing games on the fly (upon user’s request) could save storage cost, but it would take some time for the game server to be ready. In this project, the student will need to develop algorithms and simulation to empirically compare the long-term cloud storage cost when hosting on-demand games on public clouds such as Amazon EC2.Machine learning (ML) aims to construct predictive models from example input data. Single-node ML systems may not be able to cope with very large training data sets, such as ImageNet and Yahoo News Feed, which could have hundreds of millions of records. Several distributed ML systems have been proposed to reduce model training time. However, the behaviors of these systems on heterogeneous infrastructures such as public clouds have not been thoroughly investigated. In this project, we will examine the performance of popular distributed ML systems such as TensorFlow or MXNet on public cloud infrastructures, e.g., Amazon EC2, Google GCE or Windows Azure. Resource optimization techniques will then be developed to enable fast, cost-effective and convenient big model trainings on public clouds.
Assoc Prof Sun Aixin Tag based social image search
axsun@ntu.edu.sg

Social tags describe images from many aspects including the visual content observable from the images, the context and usage of images, user opinions and others. In this project, we are to study the tag-tag relationships and tag-image content relationships (i.e., the visual representativeness of tags with respect to visual concepts in social images), and their combinations in tag-based image search.

Assoc Prof Sun Aixin

Document collection topic modeling (Target for Postgraduate Student)

axsun@ntu.edu.sg

Topic modeling is to understand a large collection of documents by identifying topics formed by a group of words often occur together. In this project, we are going to study topic modeling of a given document collection base one another background document collection as knowledge base.


 

Assoc Prof Anwitaman Datta Implementing and experimenting with novel erasure codes for distributed storage systems
anwitaman@ntu.edu.sg

This project will involve understanding and implementing several families of recently proposed erasure codes, followed by thorough experimentations in computer clusters to benchmark these codes under varied workloads. The student needs to have good understanding of number theory, particularly finite fields (in order to understand the code structures), and good programming skills in general, as well as with ​socket/network programming. Knowledge of distributed systems is desirable, but not necessary.

Assoc Prof Anwitaman DattaAdapting proof of data possession techniques for efficient RAID monitoring
anwitaman@ntu.edu.sg

​Proof of data possession (PDP) techniques are usually used to detect data integrity when data is stored/outsourced with an untrusted server (e.g., on the cloud). It is a probabilistic technique and thus highly efficient, and scales well with volume of data. The purpose of the project will be to adapt the ideas from PDP in the context of RAID (redundant array of independent disks) systems. The student should have very good linux kernel + C programming skills. Knowledge of RAID systems will be an additional advantage to have.​

​Assoc Prof Wen Yonggang Social TV Analytic System: Distributed Crawler
YGWen@ntu.edu.sg

In this project, we aim to develop a big-data platform for social TV analytics. The platform consists of multiple components, including a distributed data crawler, NoSQL, various data mining algorithms, and data visualization. The candidate will work with a team of researchers and engineers to develop this system. The part of the project is to survey existing technologies and develop a distributed data crawling module.

Assoc Prof Wen Yonggang​ Social TV Analytic System: Data Visualization
YGWen@ntu.edu.sg In this project, we aim to develop a big-data platform for social TV analytics. The platform consists of multiple components, including a distributed data crawler, NoSQL, various data mining algorithms, and data visualization. The candidate will work with a team of researchers and engineers to develop this system. The part of the project is to survey existing technologies and develop a data visualization module.
 Asst/P Anupam Chattopadhyay Evolvable Video COMPression(EVoComp) system design for online video compression 
anupam@ntu.edu.sg   

With increase in the resolution of videos, higher bandwidth is required for transmission, thereby necessitating the need for compression. The project aims at developement of an FPGA based system that is – adaptive, so that it adapts according to changes in video characteristics and – on-line so that it can compress/decompress a video stream as it is being transmitted/received. The EvoComp system will allow evolutionary algorithms to change the hardware configuration  in real time to change compression methodolgy depending on video characteristics. We aim to benchmark the developed system with the existing video compression technqiues. 

Asst/P Anupam ChattopadhyayCollaborative energy-aware downloading for locally connected Android Mobile Devices
anupam@ntu.edu.sg

Energy conservation is a critical aspect of mobile devices. Services like downloading content are energy hungry. The project aims at developemnt of energy efficient techniques to manage downloading large files by collaborative downloading across locally connected (over Bluetooth or Wifi) Android devices to maximize throughput and at the same time taking into account the characteristics of the peers such as network usage, current cpu utilization, battery level, etc for partioning the download. 

Asst/P Erik Cambria

Concept-Level Sentiment Analysis with SenticNet

cambria@ntu.edu.sg

This project is in the context of sentiment analysis or opinion mining. The opportunity to capture the opinions of the general public about social events, political movements, company strategies, marketing campaigns, and product preferences has raised growing interest both within the scientific community, leading to many exciting open challenges, as well as in the business world, due to the remarkable benefits to be had from marketing and financial prediction. 

The candidate will focus on improving the current version of the Sentic API (http://sentic.net/api), a publicly available resource for concept-level sentiment analysis. Possible improvements are listed below.

 1. KNOWLEDGE EXPANSION
Enriching SenticNet with new concepts.

Possible ways to do this are:

a) write some code (preferably in python but java also can) to import concepts from other resources, e.g., Bing Liu's Opinion Lexicon, MPQA Subjectivity Lexicon, SentiWordNet, Harvard General Inquirer, LIWC, etc.
b) crowdsourcing, e.g., through surveys, quizzes or games
c) an ensemble of the above

2. CONSISTENCY CHECK
Finding errors in SenticNet, e.g., in terms of polarity or semantics. 

Possible ways to do this are:
a) do it manually
b) use the same code/resources mentioned before to find clashes (e.g., positive polarity in SenticNet but negative in SentiWordNet)
c) an ensemble of the above

3. FEATURE EXTENSION
Enriching SenticNet w/ new add-ons.

Possible ways to do this are:
a) add POS tags
b) add mood tags
c) add category tags

 Possible extra activities for the candidate include working on any of the modules of our model for concept-level sentiment analysis (http://sentic.net/clsa-model.pdf).​

​Asst Prof Lam Siew KeiReliability-Aware High-Level Synthesis for Embedded Computing Platforms
siewkei_lam@pmail.ntu.edu.sg 

​Single Event Upsets (SEU) induced faults poses a serious reliability problem in safety-critical embedded systems. Existing SEU mitigation techniques often incur unacceptable performance-power-cost overhead. The main objective of this project is to devise techniques that automatically synthesize reliable embedded systems from high-level design descriptions guided by constraints and target hardware architecture specifics. The candidate will investigate lightweight SEU mitigation techniques for fault tolerant custom hardware and realize the proposed techniques on state-of-the-art FPGA platforms. The candidate is expected to be experience in digital design and Verilog HDL, good foundation in data structures and algorithms, and good programming skills (C/C++).​

Asst Prof Lam Siew KeiAccelerating Feature Detectors for Real-time Vision-based Applications 
​siewkei_lam@pmail.ntu.edu.sg 

​Feature detection is a fundamental step in many real time applications such as video tracking, visual SLAM and robotic navigation. However, existing implementations for feature detection is highly compute intensive and becomes a bottleneck for real time vision tasks. This project aims to develop hardware-efficient feature detectors on FPGA.​

Assoc Prof Vun Chan Hua, NicholasRNS based embedded signal conversion and processing techniques
RNS enables highly efficient hardware based signal acquisition and processing techniques. This project is to investigate the optimization of the novel Residue Number System (RNS) based signal processing techniques that were invented in SCSE as listed below.

a) RNS encoding based folding ADC. ISCAS 2012: 814-817
b)
A New RNS based DA Approach for Inner Product Computation. IEEE Trans. on Circuits and Systems 60-I(8): 2139-2152 (2013)

The project involves finding the most suitable moduli set to implement the most efficient system based on the above techniques, in term of balanced moduli,  efficient reverse conversion, as well as further novel features such as error detection and correction in the acquisition and processing stages.

​Assoc Prof Anwitaman DattaWeeding the fake out from social media
Anwitaman@ntu.edu.sgSocial media plays a important role in modern lives - in how information is shared and consumed. Yet, a lot of information on such platforms are not authentic, and furthermore, irrespective of the inherent nature of the information itself, oftentimes, the extent to which it has actually penetrated the network (say in terms of likes) on the platform is incorrect. The latter (fake likes) is used for a myriad of reasons – e.g., to project a better value to a social media asset than it actually has, or to give it a critical momentum first in hope of gaining organic traction afterwards, and so on. This project will aim to study the propagation of fake metrics on social media services such as fake likes on Instagram. The purpose of the project will be to develop automated and robust solutions to detect inorganic and fake likes on Instagram using Social Network Analysis techniques, machine learning and anomaly detection on a large dataset of social media (Instagram) users. The student should have good programming skills and a good understanding of statistics and machine learning. Prior knowledge of Web 2.0 APIs is desirable, but not necessary. The project has many steps – from data gathering, cleaning to analysis, and is aimed to be for 5–6 months. 
Prof Jagath C RajapakseIdentification of cancerous skin lesions
ASJagath@ntu.edu.sg

Skin cancers are the most common human malignancies in fair skin populations. The aim of this project is use deep convolution neural networks to identify skin cancer from skin lesion images. The student will develop algorithms to implement convolution networks and classify skin lesion images into cancer and non-cancer types.

 

The aims of this project are

  1. to extract potential features from skin images for cancer classification
  2. to develop deep a deep convolutional neural network to identify cancerous skin lesions images from cancer images.

 

The student will develop deep neural networks by using Caffe software (http://caffe.berkeleyvision.org/).

Prof Jagath C Rajapakse
Classification of skin cancer from skin lesion images 
ASJagath@ntu.edu.sg

Skin cancers are the most common human malignancies in fair skin populations.  Though melanoma has the highest mortality, other non-melanoma cancers are more common. The aim of this project is to classify cancerous skin lesions into different cancer types from skin lesion images by using deep convolution neural networks.

The aims of this project are

  1. to extract potential features from skin images for cancer classification
  2. to develop deep a deep convolutional neural network to classify cancerous skin lesions into different cancer types.​

The student will develop deep neural networks by using Caffe software (http://caffe.berkeleyvision.org/).

Share Article