Automated bug assignment: Ensemble-based machine learning in large scale industrial contexts
Example 1
LAST UPDATED
27/06/2022
Example 2
LAST UPDATED
27/06/2022
Visual Abstract Activity
Paria Eskandarpour 33433526
Don’t have any quant example!
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Peter Ince 20716966
Technological rule: to explain a new exploit and identify vulnerabile smart contracts on the Algorand blockchain using our tool that we developed
Developers and users of Algorand Smart Contracts
Algorand Smart Contract was exploited and there has been no description of mechanism of action, and no way to detect the exploit in other smart contracts
Design science methodogy
Smart contracts were exploited with a new expliot
We provide a case study of mechanism of action, develop a static analysis tool tool to identify other vulnerable smart contracts, and verify that identified smart contracts are exploitale
Performed a systemic review to identify other similar works, and tested against all current smart contracts
Case study of mechanism of action, development of a new tool to identify identified exploit, analysis of current smart contracts with developed tool and confirmation of identified smart contracts vulnerabilities
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Hira Naveed (33510989)
Technological Rule: To develop human-centric machine learning systems using model driven software engineering
Systematic literature review of Model-driven engineering approaches for ML
Software systems with ML often do not consider human aspects which results in systems that do not fully meet user needs
Problem observed in several recent papers
Modeling human aspects as first class citizens with ML components and generating human centric ML systems
Model-driven engineering
Usability study with developers and end users + evaluation on industrial case study
Evaluation on industrial case study and user study
LAST UPDATED
27/06/2022
Name: Yi Li (29014271)
quantify physical exertion in AR/VR interaction using natural body gestures
Lacking of indications of physical exertion in XR applications,brings potential physical injuries to the general XR users.
A metric to help XR designers understand the physical exertion needed for the applications and guide them improve the design to avoid physical fatigue/injuries.
AR/VR Applications
Evaluated in 12 participants by comparing the existing model with the object fatigue measurements like surface EMG, using ART ANOVA.
An existing shoulder fatigue model (Consumed Endurance,Aka, CE) shows high correlation with the self-report fatigue but hasn’t been evaluated by objective approaches yet.
The XR community will be regulated based on the standardised physical exertion levels needed in the applications. XR designers will improve their designs based on this information and the general users can better protect themselves in the future XR engagement.
An existing model: CE can indicate the spent physical effort based on the shoulder gestures.
Fatigue evaluation in XR interaction.
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name:Adi Daman Huri (25416375)
Distinguish claims that are made without being factually supported (lack of) by evidences
To improve the ability discerning misinformation in claims made from Social Network Sites, we utilise explainable fact-checked evidences.
Discerning ability of humans of misinformation
Discerning test: Baseline (blind test) comparison vs Assisted forms
Discerning Misinformation through Automated Fact-Checking on Social Network Sites
To understand which aided representation is effective in discerning misinformation
Automated fact-checking with emphasis on explainability (aided representation)
Correlation analysis to which aided form is optimal
HCI: Explainable form of evidences
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Bhagya Maheshi, 33377146
Dialogic Feedback Through Responsible Inquiry: Design Recommendation for Learning Analytics
To understand how students engage in dialogic feedback in large student cohorts use feedback request forms based on responsible inquiry
Providing dialogic feedback to large student cohorts based on their needs and aligning to learning design is difficult
Responses received from 2nd year software development undergraduate students for 3 consecutive surveys on their assignments
Responses received from 2nd year software development undergraduate students for the 4th survey regarding their experience with feedback request forms
Learning Analytics based feedback tool that empower dialogic feedback based on both student and teachers’ requirements while aligning to learning theories
Develop business requirements for initiation of feedback dialogues from student feedback requests based on Social Cutural Theory, Meta Cognitivism and Social Constructivism
Higher education institutions with large student cohorts
Development of a learning analytics based feedback tool that supports dialogic feedback
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Shaozhang Dai (27143333)
Most existing methods for remote object interaction involves recalling and change of PoV, which affects the perspective of the visual content during the process.
To achieve better immersive experiences when interacting with remote object in virtual reality with a new interaction technique, Yo-Yo interaction
A new interaction technique that allows direct interaction with remote object while the PoV is constant to keep the perspective of the overview.
New interaction technique
Existing approaches for remote interaction like Go-Go (or other non-linear extension interface), teleportation, rescalling, or WIM have a common issue of large cognitive load due to the perspective changes in the process.
Experiment designed to collect quantitative data (performance measure including: accuracy and efficiency) and qualitative data (subjective preference of the new technique and the existing technique)
New interaction technique, Better performance in terms of efficiency (time cost to finish specific task) and accuracy (precision of the movement towards the target)
Related work to achieve the interaction with remote objects
New approach that can have equal/better interaction experience while keeping the perspective of the overview of the visual content
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Mya Ballin (33694761)
To improve the quality of recordkeeping in out of home care and care leaving planning through the design of novel digital record types.
The current system of out of home care (OOHC) and care leaving records does not adequately address the needs of the children they document and are intended to serve
The creation of dynamic, digital recordkeeping tools that incorporate a new governance paradigm
Most relevant to archivists and records management professionals; care leavers and care experienced children
Will compare and contrast responses of focus group/SMEs to form structure and their perceived/proposed workflow when working with the material as compared to traditional records
Will propose a new care-centred model for the possibilities of recordkeeping, offering an example of how care and love focused governance can be applied in the OOHC recordkeeping context.
Documented in existing literature
Focus groups and surveys
Recordkeeping informatics and HCI theories; institutional theory; archival studies and archival diplomatics
LAST UPDATED
27/06/2022
Visual Abstract Activity
Asma Alzahrani 32043384
Learning analytics research has begun to recognise the importance of understanding the student perspective towards the services that could be potentially offered; however, student engagement remains low. Furthermore, there has been no attempt to explore whether students can be segmented into different groups based on their expectations towards learning analytics services
The current exploratory work addresses this limitation by using the three-step approach to latent class analysis to understand whether student expectations of learning analytics services can clearly be segmented, using self report data obtained from a sample of students at an Open University in the Netherlands
An exploratory latent class analysis
The results are discussed in relation to previous work on student stakeholder perspectives, policy development, and the European General Data Protection Regulation (GDPR).
Students, higher education senior managers
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Tongguang Li (26594609)
Learners reported that they are reluctant to follow the auto-generated instructions because they felt that instructions were not tailored to their use of overall learning strategies
Adding more learning features in the current algorithm for adaptive supports
Related works discovered that learners did not comply with the instructions
Application of the updated algorithm
Problem observed in the real learning context
Learning data is collected from 200+ participants in two rounds within a year
Previous studies insufficiently addressed learners’ cognitive behaviors (i.e., adoption of learning strategies) in designing adaptive interventions
To achieve more effective instruction to promote learners’ self-regulated learning in the context of online learning use learning analytics techniques
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Jinchun Du (26848449)
Finding and computing optimal shortest obstacle-avoiding path between a start-destination pair in Euclidean space efficiently.
Proposed an algorithm named EHL to preprocess data information of a given map and retrieved it for online query computation.
Euclidean Shortest Path Problem (ESPP) is a well studied problem and has many past literatures proposing optimal and efficient solutions.
Experimented on four well known benchmarks in the domain and compared with two state-of-the-art approaches. Run time and memory was compared.
First work to adapt Hub Labeling in Euclidean space for pathfinding. The proposed algorithm outperformed state-of-the-art algorithms by 1 to 2 orders of magnitude.
To efficiently compute shortest path between any given two points in a Euclidean space using Euclidean Hub Labeling algorithm.
Ultrafast Euclidean Shortest Path Computation using Hub Labeling
Past literatures provide evidence of problem application scenario
Experimental validation of algorithm effect compared with past works
Adapting a well known strategy in road network to Euclidean space
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Lifi Huang (32263619)
Individual acoustic classification requires a substantial amount of labeled data, which in turn requires extensive amounts of expert labeling hours
Develop a classification method that requires less labeled data.
Comparison of new method to existing methods
Instead of training an end-to-end trained model, incorporate existing domain knowledge into the framework to reduce the amount of labeled data required.
Train various models/perform ablation studies and compare their respective metrics.
A major obstacle on the path to passive individual acoustic monitoring is the amount of labeled training data required to train a well-performing classifier.
Developed from interviews with biological domain experts. Also a general problem in all classification tasks/literature.
Technological rule: Reduce the amount of expert labeling time required to train an individual acoustic classifier.
Incorporate existing expert domain knowledge into the training framework.
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Yue Yang (32631170)
To understand how does different form information shape the behaviour of individual ant and therefore change the dynamics of the whole group?
We applied the ant foraging paradigm to address gaps in our project by analyzing the underlying learning mechanisms of foraging behavior and modeling individual ant behavior using reinforcement learning by characterizing foraging as a stochastic game to enable qualitative assessments beyond biological boundaries.
Achieve highly expressive mathematical models of social insect learning behaviour that are biologically plausible with reinforcement learning
The development of formalism and mathematical approaches can serve as a formal language for interpreting the mechanism of collective decision-making among enormous populations.
There has been limited research on the combination of team and individual learning for social insects, despite its potential to closely reflect real-world scenarios.
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Ming Kang (33376263)
How Does Financial Development Impact Economic Growth? To find evidence from real data.
Using the quantile linear regression model.
Empirically, economic growth responds positively to financial development when the level of financial development surpasses a threshold value.
Most previous studies predicted a linear relationship between finance and growth. However, few studies argued that financial development occurs endogenously as the economy reaches a critical threshold of economic development.
To examines the nonlinear relationship between financial development and economic growth using the threshold regression model.
Theoretical literature suggests that financial development can promote economic growth.
Several empirical studies highlighted that financial development affects economic growth through capital accumulation and technological innovations.
The modeling approach is based on endogenous growth theories and Followed aggregate production function.
How Does Financial Development Impact Economic Growth in Pakistan?: New Evidence from Threshold Model
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Chanathip Pornprasit (Oat) Student ID: 32131925
Developers have to wait for long time to get feedback of their submitted code changes from reviewers
Develop deep learning approach to help developers revise their submitted code changes
Developers at Microsoft have to wait for 15-20 hours to get feedback of their code changes from reviewers.
Evaluate in 3 open-source software hosted on Gerrit, by measuring accuracy
We leverage pre-trained code model (CodeT5) and conduct evaluation in time-wise setting.
Develop approach based on pre-trained code model
Measure accuracy on test set
Recent work showed that problem exists in real-world
To help developers get feedback of their submitted code changes faster in code review, use deep learning model to get revised version of the submitted code changes
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Wannita Takerngsaksiri Student ID: 31537162
The existing code generation works are either constrained by grammar-structure or pretrained LM, but less account for compilability.
Apply compiler feedback to the source code language modelling.
The experiments are performed on CodeSearchNet (Code-to-Code) and AdvTest (Text-to-Code) datasets.
The solution is compared with the state-of-the-art baselines (e.g., Transformers, GPT-2, CodeGPT) on the evaluation metrics of Edit Similarity and Compilation Rate
The paper utilizes RL to the code generation task with compiler feedback as a reward function.
To achieve higher compilability of the code generation in code-to-code and text-to-code tasks using three-stage pipeline including Language Model, Fine-tuning, and Reinforcement Learning techniques.
Compilable Neural Code Generation with Compiler Feedback
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Mengting Huang 31518672
To achieve better scanning result in MRI and CT using Machine Learning
Restoring clinical image that is corrupted during the acquisition
Bridging the theoretical gap of solving inverse problem using machine learning
Image
Inverse Problem
Develop theory to solve inverse problem using deep learning as a extension of traditional signal processing
Patient suffered from long time scanning and being called back for rescan due to the corrupted image
Evaluate in clinical practice, assist with radiologist review
Increase patient throughput, save time, more precise result for decision making
Theory from deep conv framelets
Retrospectively validation
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Yonghui Liu (29028779)
Individual Code detection and unification
Generate unified intermediate language for different languages used inside React Native Android app
Comparison of new method to existing methods
Enable the non-java static analysis for android app, to enable an unified static analysis framework.
Evaluate the development doc and other tools.
Android app static analysis suffered from the lack of non-java language static analysis.
Developed from review of different of tools and code used in the development of Android app
Technological rule: Enable the complete unified code analysis on React Native Android App
Incorporate static analysis framework and further new tech on them.
A unified static analysis framework for Android applications.
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Chamath Abeysinghe Student ID: 30878373
Insect tracking is a labour intensive and time consuming task in collective behaviour analysis experiments
Deep learning based automated multi object tracking, applies domain adaptation
Problem is experienced in practical lab environments.
Related work, previous experience
Applied detection based tracking, joint-detection-tracking methods with domain adaptation
Compare labelling times and tracking accuracy for 3 datasets, using 3 different approaches: manual, DL based automation, DL based automation with domain adaptation
Use domain adaptation with DL based joint-detection-and-tracking
Achieve accurate efficient multi object tracking in highly crowded environments.
Evaluate MOT metrics
LAST UPDATED
27/06/2022
Visual Abstract Activity
Student ID: 33316058
Po-Yao (Cosmos) Wang
Lucid dreaming is interesting, but it is difficult to control for beginners
Interactive technologies (TES, EMS, VR)to help people manipulate their dreams
Apply tech. to two groups and see the difference
Lucid dream beginners?
Validated by the comparison of two groups
The first paper helping people experience lucid dreams with interactive tech.
To help lucid dream beginners experience vivid lucid dream with interactive tech.
RW show that external stimuli affect dreams
Related work shows that lucid dream need practice
Lucid Reality: using interactive tech. To help lucid dreamers experience vivid lucid dreams as virtual reality platform
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Yueyi Ge (25695568)
To ensure that an in-home monitoring system is able to accurately detect and interpret users’ behaviour, use a method to detect the presence of visitor events.
If visitors are not detected, the system may generate false alarms that report abnormal behaviour changes caused by visitors.
Use unintrusive ambient motion sensors to collect user behaviour data and use unsupervised models to detect the presence of other people
Used real world datasets to detect visitors and compared the system performance with and without the detection of visitors.
Problem observed in real projects and evaluated on real world datasets: [Sahani et al., 2015, Balla and Jadhao, 2018, Guravaiah et al., 2022, Hu et al., 2017, Howedi et al., 2022]
Use unintrusive ambient motion sensors to collect data and use unsupervised methods to detect visitor events.
LAST UPDATED
27/06/2022
Name: Sashini 31196594
Retrospective, cross-sectional analysis
Portal enrollment and use are generally lower among racial/
ethnic minority patient populations
Mobile device use may represent an
opportunity for healthcare organizations to further engage
black and Hispanic enrollees in online patient portal use.
mobile
applications and mobile browsers may help reduce differences in
patient portal use among racial/ethnic minorities
Using data from 318,700 adults enrolled in an
integrated delivery system between December 2012 and
November 2013,
racial/ethnic minority enrollees
were less likely to access the online patient portal
(Racial/ethnic variation in devices used to access patient portals)
This cross-sectional study drew from 5 enrollee data sources: administrative data for individual demographic information, Web server logs to identify devices used to access functions, EHR records,
Patient portal engagement with minority populations
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Vidushani Dhanawansa
(27376796)
Available studies on EEG data while participants are under general anesthesia, insufficiently investigate the probability of waking up, based on the EEG data prior to a noxious stimulus.
Prediction of the alpha levels of EEG post a noxious stimulus, based on pre stimulus data, taking into account key electrodes of the EEG data.
To predict participant response post a noxious stimulus, using deep learning prediction algorithms
Enable clinical anaesthesiologists to better adjust the levels of patient anaesthesia to prevent a positive response to a noxious stimulus, whilst under surgery
Use of a systemic literature review, comparison against labelled datasets, and feedback from clinical anaesthesiologists.
Past work in this domain conduct limited investigation on predictions following a noxious stimulus, in order to be able to tune the level of anaesthesia
Based on feedback from clinical anaesthesiologists, and gaps in literature in the domain
Comparison of obtained against expected predictions on labelled dataset
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Piyumi Rameshka - 32277768
Related work discretize the dihedral angle space
Accurate sidechain dihedral angle prediction is crucial for protein structure determination. The limitations in current rotamer libraries effect many downstream applications in proteomics.
Model dihedral angle distribution in the continuous space with the aid of statistical mixture modelling and MML inference method
Technological Rule: To accurately predict the sidechain dihedral angles of amino acids in proteins given the backbone dihedral angles use statistical mixture modelling
Protein structure determination
Validate the accuracy of predictions with experimental protein structure data
Model the dihedral angle space in continuous space, use MML inference method and inference process is done unsupervised.
comparison of the prediction results with the current state of the method.
MML and statistical mixture modelling
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Tonggua
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Yue Zhang 30976316
Find optimal shortest collision free path of a set of agents while considering turning costs
Study the solution quality between considering/ignoring turning costs.
Model problem and developing algorithms to solve the problem.
Multi-Agent Path Finding (MAPF) iis the well-studied problem of planning paths for multiple agents so that they can travel from their predefined start locations to their goal locations without any collisions. This problem is relevant to MAPF while considering more realistic settings.
Experiments are from public MAPF benchmarks, the experiments settings follow a standard experimental settings in this domain. Comparison between solvers are from the existing state-of-the-art solvers and ours.
We first shows that, from all domains, direct planning with turn actions yields significantly improved solution quality compared to approaches that leave turning considerations to post-execution planning. We also shows that our solver in MAPF with turnings results in better runtime (around 1.5 - 2 times speed ups) and more instances solved.
Efficient multi agent path finding with turn actions
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Jiuzhou Han (32896565)
Technological rule: To enhance the structure awareness of PLMs in NLP using several graph masking strategies to inject local and global awareness of the input structure.
PLMs are typically pre-trained on free text which introduces domain mismatch between pre-training and downstream G2T generation tasks.
We propose graph masking pre-training strategies that neither require supervision signals nor adjust the architecture of the underlying pre-trained encoder-decoder model.
Previous work indicates that the linearisation step used in PLMs ignores the structural information of the graph (Wang et al., 2021), while explicitly modelling structured data could also lead to catastrophic forgetting of distributional knowledge (Ribeiro et al., 2021).
We evaluate our methods on three datasets: WebNLG+2020, DART, EventNarrative. We use the evaluation metrics: BLEU, METEOR, TER, BERTScore.
Our empirical findings highlight that our self-supervised strategies significantly outperform a strong underlying T5 baseline and achieve two new SotA results on two of the datasets WebNLG+2020 (Zhou and Lampouras, 2020) and EventNarrative (Colas et al., 2021). Additionally, we show our pre-training strategies are very efficient in utilising data and have a great potential for low-resource setting.
Self-supervised Graph Masking Pre-training for Graph-to-Text Generation
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Fan Yang 27596923
In practice, the interpretability of a machine learning model requires that the model has both a high accuracy and a high comprehensibility. However, the existing algorithms for interpretable machine learning models cannot guarantee their comprehensibility.
Orthogonal Gradient Boosting for Interpretable Additive Rule Ensembles
adopting a new objective function and a weight correction step in each boosting round to maximise the predictive gain per added components.
With a limited size of components in the model, we need to examine if the new approach produces higher accuracy.
We introduce a new objective function which takes the weight correction step into consideration when adding new components into the model.
We use empirical risks to measure the accuracy of the model produced by different algorithms, including the proposed method in this research. We also proposed a method to compare the balance between accuracy and interpretability.
There have been a few approaches to produce interpretable machine learning models, but if they do not limit the size of the model, the produced models cannot be understand by users.
Analyse the properties of the objective functions and see how to solve it efficiently.
To get a balance between the accuracy and interpretability of machine learning models.
LAST UPDATED
27/06/2022
Name: Ayodeji Ladeinde 28679288
To understand the use of Knowledge graph effectiveness for requirements analysis
It is not known how, or to what extent the use of knowledge graphs for requirements analysis help influence developers' perspectives on users needs.
Problem Instance
The purpose of this study is to examine how and to what extent the use of knowledge graphs for requirements analysis help influence developers' perspectives on users needs.
Purpose statement
Relevance: Requirements analysis for better software design by developers
Novelty: knowledge graph representation of entities and concepts within a set of requirements
Rigor: Use of requirements documents from diverse sources
Conduct user studies and performance metric of our approach
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name:
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name:
Saranya
Technological Rule : To generate human readable test cases
Lack of readable testcases from the existing tools(Evosuite&Randoop)
Generate test cases similar to human-written with test method names that best describes the functionality of the test Assessment of the readability of the generated tests
Relevance : A major weakness and criticism of these approaches is the unsatisfactory code quality and understandability of the generated test cases.
Rigor : Evaluate Defects4j projects and validate the correctness of test cases
Novelty : Generate test case comparable to human generated test cases
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Poobalan
To assess the maturity of big data analytics
Develop a maturity model and assessment tool
Lack of maturity model for public sector domain
maturity model for public sector domain based on a theoretical framework
Survey method with expert validation
To assess the maturity level of big data analytics implementation in public sector agencies through the use of a maturity model
Literature review, expert review, statistical analysis
Follow 6-step model development method by Becker et. al. (2009) -
Assessment of BDA Implementation Maturity in Public Sector Agencies
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Goi Yue Tian
Cryptanalysis using Deep Learning
Deep learning is in black-box
Build an Improved cypher?
a
a
Improved cypher? Not sure
Technological rule: To achieve better performance in cypher use improved cypher
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Wishma Samaraweera
Predicting the headpose of an image based on the facial landmarks
Use of GCN to decide the landmarks and the pose of the head
Will be a great help in analyzing occluded images, videos etc.
To
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Tan Pei Sze (33504768)
Title: Gender Stereotyping Impact in Facial Expression Recognition
FER datasets,their representation in the individual label is not, embedding social stereotypes into the datasets and generating a potential for harm.
Measure the discrepancy between the performance of the models trained on these datasets for the apparent gender groups.
Deuschel et al. employ intentionally biased datasets, to study the impact of these biases on the detection of action units, a problem closely related to FER.
Suggest a safety range for stereotypical bias in a dataset that does not appear to produce stereotypical bias in the resulting model.
Consider other type of biases to balance the fairness of model
Fairness in ML model
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Hoh Kar Ee
Slow sampling speed in diffusion models for image generation
Reduce sampling speed in diffusion model by proposing an approach - progress distillation
To achieve a fast sampling in image generation while maintaining the image sample quality
Score based method to evaluate the model quality
Propose a procedure known as progress distillation to reduce the sampling time
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Limei Liu(33468990)
Predict the traffic flow accurately in California
Integrate the data by week, day, and hour.
Attention mechanism
Public datasets in California
Self-attention and periodically integrate the data by hour/day/week
Consider the period regulation in traffic flow
GCN，self-attention
Attention Based Spatial-Temporal Graph Convolutional Networks for Traffic Flow Forecasting
Data and model are important
Construct a Spatial and temporal model, and select datasets
Try test on different areas in california
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Xia Zhou 31939163 ☆
Lack of theory for individual behaviors’ response to different incentives
Using knowledge from control theories and state estimation
Users’ quantitative response to a certain incentives can be simulated and analysed.
Behavior function and simulated behavior response will be included. Random utility choice model is used to simulate the choice process.
Paper title: Personalized incentives for promoting sustainable travel behaviors
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Yang Hong (32610157)
The developers need a long time to wait the feedback from reviewers in the code review.
Build an approach that can automatically recommend code review comments
Code reivew comments recommendation tasks was explored by deep learning approaches. Developers needs 15-64 hours to receive the feedback from reviewers.
Build an information retrieval approach based on a large scale of dataset from 6,388 gerrit projects and 4,901 github projects.
A simple but fast approach then previous deep learning approach for recommending code review comments.
To achieve a simple, fast and accurate approach to recommend code review comments in code review process.
Reviewers don’t have enough time to perform a detailed code review
Measure the number of code review comment that correctly recommended
Develop an information retrieval approach for recommending code review comments
CommentFinder: A Simpler, Faster, More Accurate Code Review Comments Recommendation
LAST UPDATED
27/06/2022
Most AI models are black boxes. This hinders their adoption in the real world
Explain black-box models by calculating the importance of features with respect to a specific model output
AI users who do not trust AI tools or systems in real-world scenarios because they cannot understand how the system reached a specific decision
Conduct user studies (simulated and real) that quantify user trust in the models and user performance in tasks with and without explanations. Results show that users better trust the model and have better performance when explanations are presented.
Student: Jair Ferreira
Student ID: 31090311
Generating explanations in the form of feature importance for any black-box models without depending on how the model works internally.
Why Should I Trust you? Explaining the Predictions of Any Classifier
https://dl.acm.org/doi/10.1145/2939672.2939778
To improve user trust and performance in tasks supported by black-box AI tools and systems using feature importance explanations regarding the black-box model.
Now, It’s Your Turn! http://tiny.cc/MDP-VisualAbstract
Synthesizing medical imaging modalities to generate missing data from existing modalities
Automating translate between medical image modalities
Medical Image generation and evaluating quality of synthesized image modalities which is capable of using for downstream health applications
Synthesizing of missing medical modalities from existing modalities.
Developing a model to synthesizing medical modalities
Overcoming the problem with limited modalities and health risk of the image acquisition process.
Developing model to synthesizing medical image modalities
Name: Sanuwani Hewa Munasinghe
33449201
Visual Abstract Activity
Name: Leizhen Wang
Hindsight Experience Replay
Sparse reward in reinforcement learning. RL always need engineer a reward function to make the learning process more efficiently.
Replay sampled trajectories and set some states as goals to allow sample-efficient learning from rewards which are sparse and binary
and therefore avoid the need for complicated reward engineering
A common challenge, especially for robotics, is the need to engineer a reward function
that not only reflects the task at hand but is also carefully shaped (Ng et al., 1999) to guide the
policy optimization.
Ablation experiment on 3 different robotics control tasks and deployment on a physical robot
Fast convergence and better cumulative rewards.
LAST UPDATED
27/06/2022
Visual Abstract Activity
Name: Yuanzhe Zhang 32732791 ☆
To reduce expensive cross-shard transactions in sharded blockchains
Use graph clustering to put frequently-interacted accounts togation
Users’ quantitative response to a certain incentives can be simulated and analysed.
Behavior function and simulated behavior response will be included. Random utility choice model is used to simulate the choice process.
Paper title: Personalized incentives for promoting sustainable travel behaviors
LAST UPDATED
27/06/2022