Code generation is the task of parsing natural language statements into source code like Python. In this work, we propose a re-rank based framework which aims at re-ranking these hypothesis candidates. A classifier is trained to re-rank candidates using neural based and hand-crafted features. While our re-rank framework can be easily generalized to other code generation models or natural language related tasks such as machine translation, we extend TranX with our framework and demonstrate its effectiveness on CoNaLa and Django dataset. Experiments show that our LeToRr framework improves the SOTA performance by +3.4 ACC and +4.6 BLEU on Django dataset.

We parallelized an existing automatic number plate recognition system in CUDA on GPUs and in OpenMP on multi-core CPUs, compared the performance of two implementations and finally integrated two implementations into one heterogeneous system. Finally, we have achieved on average 1.846× speedup on multi-core CPUs, 3.3351× on GPUs, and 4.1527× on final integrated heterogeneous system.

The main task of this project is about the so called Extreme Classification, where we need to deal with multi-class involving an extremely large number of labels. Specifically, the training set of Cdiscount dataset contains 12,118,359 images (6,924,452 products), and in total, there are 5,270 categories. The final goal of the project is to correctly predict products into these 5,270 categories. First, we experimented with several deep neural network models, including ResNet50, Se-ResNet50, ResNet101, InceptionV3, Xception, etc. with a limited training epochs, to test single model performance and set the baseline for further experiments. Then, we implemented Test Time Augmentation (TTA) and network ensembles to make whole framework more stable in testing stage. Then, in order to eliminate negative effect from noise data in prediction stage, we proposed a noise robust prediction aggregation algorithm. Finally, we used pseudo labeling to further fine-tune the trained neural network in a semi-supervised fashion, hoping to further generalize the model to the testing set and improve the model performance on product-wise prediction (i.e. not simply image-wise prediction). The proposed framework turns out to be more stable and effective than single base model.

In this project, we present a novel ensemble framework to extract highly discriminative feature representation of image and its application for group-level happiness intensity prediction in wild. In order to generate enough diversity of decisions, we train n convolutional neural networks by bootstrapping the training set and extract n features for each image from them. A recurrent network is then used to remember which network extracts better feature and generate the final feature representation for one individual image. We use several group emotion models (GEM) to aggregate face features in a group and use fine-tuned support vector regression (SVR) to get the final results. Through extensive experiments, the great effectiveness of our Random Recurrent Deep Ensembles (RRDE) are demonstrated in both structural and decisional ways. Our best result yields a 0.55 root-mean-square error (RMSE) on validation set of HAPPEI dataset, significantly better than the baseline of 0.78. Meanwhile, I also build an online RESTful API called GREP (GRoup Emotion Parser) and a website for demonstration purpose.

This paper summarises some of the best practices learned from an extended software engineering project completed through a collaboration of multidisciplinary faculty and several teams of computer science students. The collaboration delivered an advanced multimedia note-taking application, as an open educational resource (OER), capable of supporting both students and research into note-making practices. The project lasted beyond a single academic year, thus enabling multiple student cohort participation, and took place in an English medium of instruction, Sino-foreign university in China. The experiences and reflections surrounding the project were examined, with a number of resulting ideas for best practices.

This paper reports on a 3-year project involving three computer science student teams from different cohorts and a multi-disciplinary team of academics who collaborated on the development and deployment of an open educational resource (OER): a novel multimedia note-taking application. The project took place at an English medium of instruction (EMI) Sino-foreign higher education institution (SFHEI) in China, involving students from China, Mauritius and Indonesia — with varying levels of English and Chinese fluency — and other faculty stakeholders, all of European extraction. The academic participants came from the different disciplines of computer science, humanities, and English language teaching. A number of challenges, including language- and culture-specific assumptions, and institutional policies, were faced, and overcome. This paper summarises the background to the project, analyses the challenges encountered, and explores some of the identified best practices for collaborations between students and teachers. ​

We present a novel redundant event filtering system based on the dense word embedding called word2vec scheme incorporated with the recent advance on distributed word mover’s distance metric. We propose that our system could efficiently and effectively identify seen events out from the experimental data that could compete or even outperform against the baselines across a range of four popular bench-marking data sets. The primary contribution of this paper is the effective usage of a novel contextual distance metric on the event detection field for the purpose of alleviating semantic variation problem.

This article proposes to study the effectiveness of the recently advances on the character-level convolutional neural network (CNN) for the long-standing open challenge about the redun- dancies detection problems. We compared our recommended CNN framework with the four competitive text classification algorithms including word-level CNN, gated recurrent unit (GRU), linear support vector machine (LinearSVM), and the word mover’s distance based k-nearest neighbouring (KNN- WMD). The systematic evaluation on both Google-news and twenty-news-groups indicated that our method could outper- form the second best algorithm by at least 4.80% and 1.30% on the Google-news and twenty-news groups benchmark data respectively in terms of the standard F1 score (aka F-measure or balanced F-score).