Current Projects

• Feature Extraction and Prediction of Combined Text and Survey Data using Two-Staged Modeling || Asif Ahmed Neloy & Dr. Maxime Turgeon || August 2021 - Present
  • Poster
  • Accepted to be published on SENTIRE Workshop, 22th IEEE International Conference on Data Mining (ICDM)
  • Classification and text representation are important tasks in many Natural Language Processing (NLP) applications. Traditional NLP problems classify text, represent summarized dictionaries, and segment larger textual context. However, due to a lack of a modeling approach, their ability to extract features is limited. Moreover, the use of these methods on smaller datasets is unexplored and underdeveloped compared to more popular research areas. In this work, we introduce a two-stage modeling approach to combine classical statistical analysis with NLP problems in a real-world dataset. With 96.69% training and 70.56% testing accuracy, our proposed Convolutional neural network (CNN) and Bidirectional Recurrent Neural Networks (Bi-RNN) show exceptional performance in small dataset training settings compared to complex state-of-the-art models. Moreover, the experimental results from state-one modeling with a stacked ensemble classifier and hypothesis testing support our novel approach to combining the modeling approach.

• Disentangled Conditional Variational Autoencoder for Unsupervised Anomaly Detection || Asif Ahmed Neloy & Dr. Maxime Turgeon || January 2022 - Present
  • GitHub
  • Under Review, The Eleventh International Conference on Learning Representations (ICLR)
  • Recently, generative models have shown promising performance in anomaly de- tection tasks. Specifically, autoencoders learn representations of high-dimensional data, and their reconstruction ability can be used to assess whether a new in- stance is likely to be anomalous. However, the primary challenge of unsuper- vised anomaly detection (UAD) is in learning appropriate disentangled features and avoiding information loss, while incorporating known sources of variation to improve the reconstruction. In this paper, we propose a novel architecture of generative autoencoder by combining the frameworks of β-VAE, conditional vari- ational autoencoder (CVAE), and the principle of total correlation (TC). We show that our architecture improves the disentanglement of latent features, optimizes TC loss more efficiently, and improves the ability to detect anomalies in an unsu- pervised manner with respect to high-dimensional instances, such as in imaging datasets. Through both qualitative and quantitative experiments on several bench- mark datasets, we demonstrate that our proposed method excels in terms of both anomaly detection and capturing disentangled features. Our analysis underlines the importance of learning disentangled features for UAD tasks.

• A Comprehensive Study of Auto‑Encoders for Anomaly Detection: Efficiency and Trade‑Off || Asif Ahmed Neloy & Dr. Maxime Turgeon || January 2021 - Present
  • GitHub
  • Under Review, IEEE Transactions on Knowledge and Data Engineering
  • Unsupervised Anomaly detection (UAD) is a significantly diverse research area explored for different application domains. Over the years, many anomaly detection techniques such as clustering, generative, or variational inference-based methods are developed to solve certain deficiencies and reform state-of-the-art techniques. However, deep learning and generative models enabled a critical direction to identify unique problems and yield advanced approaches in recent years. Auto-encoder (AE) is one of the most powerful techniques that combines generative and probabilistic variational modeling with deep architecture. The fundamental idea in auto-encoders (AEs) is to learn the distribution of data in such a way that consequential sample data can be generated. Data generation and the concept of adopting generative modeling have led to enormous research and variations in the design of AEs in the field of unsupervised representation learning. In this study, we reviewed 11 AE divided into three categories and conducted experiments to differentiate their reconstruction ability, sample generation, latent space visualization, and the accuracy of classifying anomalous data. While experimenting, we also carefully observed the scope of reproducibility with different training parameters. Fashion-MNIST and MNIST datasets are used in the reproducibility process to generate results from similar model setups and hyperparameters. Finally, through the results obtained over the experiments, we undertake to riddle out the efficiency and trade-offs among AEs.

Past Projects

• Predicting hypertension using machine learning models || Asif Ahmed Neloy & Dr. Li Xing || VADA program Internship || June 2022 - September 2022