Dates: 1st Aug - 1st Sep 2021

Description:

• Data source: EEG raw data collected from human brain
• This is the eight channel EEG raw data collected from human brain under controlled environment having no internal or external noise. The data was collected from COMSATS university Islamabad. Data is composed of 90 % students and 10 % teachers.
• For every subject, two set of data was made. First set of data was taken when the subject had eyes open, and second set was taken when eyes of the subject were closed. The hardware collecting the data was CYTON 8th channel device.

Impact:

• Received 3 non-novice upvotes and 150+ views from students at the COMSATS university Islamabad.
• Contributed to the learning of applying ML models to classify EEG brainwave and evaluate results with concise comments throughout.
• Provided a prototype with the potential of deployment and real-world usage in mind.
• Provided experiment plan and results with a variety of signal processing techniques to be used as a foundation for further investigation.

Reference:

• Link to the Kaggle Project

Dates: 3rd June - 16th Aug 2019

The University of Edinburgh - MSc AI: Final Project

Supervisor: Prof. Stuart Anderson(Deputy Head of Informatics Department)

A WeChat Mini-Program has been developed to collect, visualise and calculate the preferences and opinions of a group of decision makers and support consensus.

UML, MVC design pattern, functional testing and GUI testing, JavaScript, WXML, WXSS were used for implementation. Quasi-experiment has been conducted and evaluated qualitatively and quantitatively. Decision theory has been incorporated in the design.

Date: 10th June 2019 - 30th Aug 2019

Side Project: NLP, Natural Language Generation

Researchers: Kleber noel and Chon In (Haydn) Cheong

Web crawling and scarping were used to extract 1 Million lines of lyrics from LyricsMaster and LyricsGenius. With aggregation following standardization. Data cleaning was done by using cross-entropy, word frequency and RegEx, etc. Data is then feed into a Seq2Seq model to learn. BERT can potentially be used for future work.

Dates: 15th Feb - 5th Apr 2019

The University of Edinburgh - MSc AI: Data Mining and Exploration

Authors: Cecilia Cobos Santes, Chon In (Haydn) Cheong, Hong Tin Chan, Ivaylo Genev

Supervisor: Dr.Arno Onken

Motivation: In the ever-growing internet advertising system, it is considerably beneficial foradvertisers to promote image-based advertisements. However, such images canincrease a page’s load time, negatively affecting users’ browsing experience. Therefore, the desire arises to remove advertising images.

A number of different ways for exploratory data analyses (EDA) were presented, intended to inform data pre-processing steps and modelling. As a result of highly-imbalanced classes and somemissing data, methods such as SMOTE and multiple imputation are employed to ensure reliable classification. Moreover, three classification-based models foradvertising detection are proposed. It is found that with the combination of randomforest for missing data imputation, LDA for dimensionality reduction, SMOTEfor imbalanced dataset, and multi-layer perception (MLP) classifier, over 99%balanced accuracy can be obtained!

14th Feb - 22th Mar 2019

The University of Edinburgh - MSc AI: Machine Learning Practical

10 technical indicators were incorporated into a deep neural network as the baseline. Sentiment analysis based on news headlines, google trend and attention mechanism were incorporated to determine their effects on the predictive power of the model.

Feb - Mar 2019

The University of Edinburgh - MSc AI: Natural Language Understanding, Generation and Machine Translation

Advisor: Dr Adam Lopez

A seq2seq model with bi-directional LSTMs was used as the baseline. Data-preprocessing was done by refering to correlation coefficient between source and target lagnauges (English and Japanese in this case). Data visualisation was used to inspect the distribution of the sentence lengths to spot outliers and compare their lexical variety(type-to-token ratio) and mark rare words with UNK symbol.

An extended models with enhanced decoding layers were implemented, by replacing greedy decoding with beam search with brevity penalty to address short sentence problem. An architecture with more layers was also implemented with an aim to boost performence. Effects on model quality were examined based on training loss, validation loss, perplexity and BLEU scores. Attention mechanism was then augmented to improve alignments learned by the model. Qualitative and quatitative appraoch were used in the evaluation section.

Jan - Feb 2019

The University of Edinburgh - MSc AI: Natural Language Understanding, Generation and Machine Translation

Advisor: Dr Adam Lopez

After hyperparameter tuning for the recurrent neural network language model(RNNLM) based on cross-entropy, it was used to predict subject-verb agreement over sentences in a corpus. Accuracy of 74.30% and 73.15% were recorded for developent set(1,000 samples) and test set(4,000 samples) respectively.

Two hypothesis were then tested: (1)Does the prediction accuracy of subject-verb agreement negatively correlated with the time steps between the subject and verb? (2)By comparing the normalized loss results from Q2 and Q3, which differ in the prediction objective, would a clear effect of different hyperparameter values be shown? A clear trend of significant increased in perplexity and mean loss against time steps was found. This implies the potential of increased performance by cleaning our data before making predictions, by excepting input with a limited value of time steps from head to predictive target.

Future work was then stated: If the number time steps between subjects and verb of input sentences has a direct effect on accuracy, what is the limit of time steps in between the subject and verb for our pre-trained model to stay at an accuracy above 50%? This question has been raised since a system with prediction accuracy lower than 50% performs worse than a random guess.

8th Feb - 26th Feb 2019

The University of Edinburgh - MSc AI: Data Mining and Exploration

Poster authors and presenters: Hong Tin Chan and Haydn (Chon In) Cheong

Paper authors: Erich Schubert, Michael Weiler and Arthur Zimek

Marker: Dr.Arno Onken

This poster and presentation aim to critically review a published research paper - "Outlier Detection and Trend Detection: Two sides of the same coin". This study aims to examine the relationship between different outlier/anomaly detection methods.

We condense this paper into an A0-size poster and summarised the paper in a 10 minutes presentation, and assessed by a Q&A with our lecturer - Dr.Arno Onken.

5th Nov - 26th Nov 2018

The University of Edinburgh - MSc AI: Accelerated Natural Language Processing

Marker: Prof. Sharon Goldwater

Given a dataset from Twitter, a variety of similarity measures, namely Consine, Jaccard, Dice, Euclidean and Manhatten, were implemented and compared to investigate their effectiveness to distributional semantics of words and also sentiment analysis of sentences. After preprocessing such as removal of stopwords, repeated letters and rare words, NLTK library, word2vec were used. The relationships between words, like synonyms, antonyms or hyponyms were also investigated.

The sensitivity of each similarity measures to word frequency, systematic difference between measures, special cases were also considered and analyzed.

5th Nov - 23rd Nov 2018

The University of Edinburgh - MSc AI: Machine Learning Practical

This focus on the implementation of CNNs and a variety of pooling methods. Investigated the effects of different architectures on the performence of the CNN.

A research paper was written, after a comparison of different dimension reduction methods, results have shown that average pooling method is superior to the max pooling, striding and dilation in terms of accuracy in this setting.

5th Nov - 23rd Nov 2018

The University of Edinburgh - MSc AI: Machine Learning Practical

This focus on the implementation of CNNs and a variety of pooling methods. Investigated the effects of different architectures on the performence of the CNN.

A research paper was written, after a comparison of different dimension reduction methods, results have shown that average pooling method is superior to the max pooling, striding and dilation in terms of accuracy in this setting.

24th Oct - 18th Nov 2018

The University of Edinburgh - MSc AI: Machine Learning and Pattern Recognition

Marker and Instructor: Dr Iain Murray

A dataset from the UCI machine learning repository was used in this project. Some features have been extracted from slices of CT medical scans. This was a regression tasks of predicting the location of a slice of CT medical scans. After spliting the data into training, validation and test sets, data pre-processing was done by removing constant features and outliers. Linear regression with regularization was set as the baseline of the experiment. PCA was used to decrease the input dimensionality, since the baseline overfit when there are a lot of features. Extra binary features were introduced to make the input data more salience. To simplify the problem, binary classification tasks were invented to fit the data, so that there is no need to think hard about how the values might be distributed and how we should process them. 10 threshoded classification problems were framed for fitting the data. Given a feature vector, we can now obtain 10 different probabilities, by 10 logistic regression models.

A small neural network was also used to fit the data. A least square cost function and gradients for this neural network was implemented. Training was done after hyperparameter tuning. Its test accuracy was compared with the baseline and the baseline was extended by deploying basis functions.

20th Oct 2017 - 11th May 2018

The University of Sheffield - BEng EEE Final Project

Primary supervisor: Prof. John David; Secondary supervisor: Prof. Shankar Madathil

An avalanche photo-detector(APD) with material AlAsSb for photon absorption layer was fabricated. Characterized by the sensitivity, IV curve, CV curve and frequency range. The results of the AlAsSb with InP as substrate layer was compared against a banchmark with Si as the detection layer. The results have shown that AlAsSb is a promising material for rapid optic-dedection.

5th Nov - 23rd Nov 2018

The University of Edinburgh - MSc AI: Machine Learning Practical

This focus on the implementation of CNNs and a variety of pooling methods. Investigated the effects of different architectures on the performence of the CNN.

A research paper was written, after a comparison of different dimension reduction methods, results have shown that average pooling method is superior to the max pooling, striding and dilation in terms of accuracy in this setting.

6th Dec 2017

The University of Sheffield - BEng EEE Final year

The title was granded by National Instruments after LabVIEW Core 1 and 2 training with more than 3 months of development experience, and passing their official exam. This certificate indicates a broad working knowledge of the LabVIEW environment, a basic understanding of coding and documentation best practices, and the ability to read and interpret existing code.

For further details about CLAD: https://sine.ni.com/nips/cds/view/p/lang/en/nid/14438

21st Oct - 9th Dec 2017

The University of Sheffield - BEng EEE 2nd year: Interdisciplinary

Supervisor: Prof. Maria Merlyne De Souza

We delivered solutions to the Advanced Manufacturing Research Centre (AMRC) representatives. The solutions aim to monitor the temperature and pressure of a composite material in real-time within a hot-press machine.

3rd Feb - 7th Feb 2017

The University of Sheffield - BEng EEE 2nd year: Interdisciplinary

Worked in a multidisciplinary team to design a ski-helmet with a goggle, which has built-in augmented reality(AR) for pathfinding. This was a short project which focus on the marketing, intellectual property and planning aspects of an engineering project.

Outcome: We designed the AR goggle as a product, taken the patent aspects into considerations and created a business plan. An elevator pitch and Q&A session were delievered to project managers from an engineering company.

1st Oct - 17th Oct 2016

The University of Sheffield - BEng EEE 2nd year: Coursework Module

A remote toy car was built. It has a mechanical arm to capture a small object and travel at high speed. It was used in an internal robotic football competition and proceeded to quarter-final.