Asif Khan

4.31, Informatics Forum, University of Edinburgh· mak4086@gmail.com· asif.khan@ed.ac.uk

I am a Data Science and Artificial Intelligence, Ph.D. student at the University of Edinburgh. I am a part of BayesWatch research group where I am supervised by Prof. Amos Storkey. I am interested in developing topology and geometry aware machine learning methods to understand the underlying structure of data.


Experience

Master Thesis Student

Sony, Stuttgart

Thesis:Generative Adversarial Networks for Unsupervised Cross-Domain Speech-to-Speech Synthesis
The idea of unsupervised image-to-image translation has been extensively explored for style transfer with an application like translating a photograph to the “Monet Impression, Sunrise”. In our work, we build on this idea and extend it to develop a speech-to-speech synthesis framework. We define speech-to-speech synthesis as the task of translating the voice of the male speaker to that of the female speaker and vice versa. We work with the time-frequency (TF) representation of speech. We further present the key challenge associated with using the TF representation in the training of neural networks. Finally, we present our solution to a problem using the consistency condition of TF as an additional constraint in the optimization problem.

Mar 2019 - Aug 2019

Student Assistant

Data Science and Data Engineering group, BIT, University of Bonn, Germany

Worked on spectral methods for network alignment.

Sep 2018 - Sep 2019

Student Assistant

SDA, University of Bonn, Germany

Worked on developing machine learning methods for knowledge graph analysis.

Oct 2017 - Aug 2018

Research Assistant

CBRC,KAUST,Saudi Arabia

Worked at the intersection of Artificial Intelligence & Bioinformatics with focus on developing Machine Learning solutions for Life Science problems. Projects, I worked on:
· Protein function prediction
· Multi-modal learning for disease-gene prioritization
· Deep Learning model to predict plants traits from raw Images

Jan 2016 - May 2017

Research Intern

Rapid Rich Object SEarch Lab, NTU, Singapore

Fine-grained object classification : Worked with deep CNNs to prevent the classification of an object to its visually similar class with focus on dataset of visually similar handbags.

Jan 2016 - May 2017

Research Intern

Cybersecurity Education & Research Centre, IIIT Delhi New Delhi, India

Source camera identification : Worked with probabilistic methods for modelling noise distribution of cameras and use it with a manifold-based learning to identify the source camera of an image.
Image Forensic : Worked on an algorithm for detection of double compressed JPEG images using gaussian mixture model (GMM) and support vector machine (SVM).

May 2014 - Jul 2014, Dec 2014 - Jan 2015

Education

University of Bonn

Master of Science
Computer Science

GPA: 1.1/1.0 (1:Best,5:Worst)

Oct 2017 - Sep 2019

LNM Institute of Information Technology

Bachelor of Technology
Electronics and Communication

GPA: 8.94/10

Jul 2012 - May 2016

Tutoring

University of Bonn

Knowledge Graph Analysis
WS17/18, WS18/19

LNM Institute of Information Technology

C Programming Lab
· Fall13
Digital Signal Processing Lab
· Fall15

Skills

Python · C · MATLAB · SQL · SPARQL · Pytorch · Tensorflow · NNabla · Sklearn

Projects

  • Soccer ball localization
  • In this work, we address the problem of soccer ball localization using deep convolutional neural networks (CNN). We first identify the problem of localization in an image and solve it using variants of fully convolutional networks (FCN) referred as SweatyNet. Later we extend it and use set of sequences of images to further improve the task of localization. We use temporal models for this purpose namely ConvLSTM and TCN on the top of SweatyNet. Instead of training from scratch, we take advantage of transfer learning and finetune SweatyNet. We evaluate all our experiments on the novel dataset of 4562 images created as a part of cudavision lab. Furthermore, we present empirical results to support the effectiveness of deep CNN for the soccer ball localization task and show improvement with transfer learning approaches when using temporal sequences of images.

  • Game Artificial Intelligence
  • In this work we identify three AI problems associated with game design: a) strategies for turn based games, b) game trees & path planning and c) behavior programming. We present our solution to the problems using various search and machine learning methods discussed in the lectures of Game AI at the University of Bonn. We present our solution with an enriched analysis of the computational challenges.

  • Commonsense Knowledge Base Reasoning
  • In this work we focus on commonsense reasoning. We do so by formulating a problem of knowledge base completion on ConceptNet. Such a formulation could help in increasing the coverage of knowledge in ConceptNet. Recently a lot of methods in knowledge base completion have been developed which can infer new facts based on existing knowledge. This work implements some such methods and extends them for commonsense knowledge base a.k.a ConceptNet. Doing so helps in learning common sense enriched word embeddings which can be used for different tasks like text classification, sentiment analysis, entity resolution, question answering etc.

  • Multi-Modal Learning for Disease-Gene Prioritization
  • The identification of genes associated with diseases plays a vital role in improving medical care and in a better understanding of gene functions, interactions, and pathways. In this work, a new method is developed to jointly learn representations of disease and gene entities from text-mining and biological knowledge graphs. These representations are used to discover novel associations for the disease.

  • Learning Representation of Biological Knowledge Graph
  • The Semantic Web has been highly successful particularly in biology and biomedicine. The main success of the Semantic Web in the life sciences has been the development and use of ontologies. While the Semantic Web provides a useful way to represent and integrate biological data, networks, and background knowledge provided by ontologies, its use is often limited to retrieval and not to validate and infer new biological relations, or analyze a novel data set. There is also a certain lack of machine learning methods, particularly in biology, especially to use data together with ontologies. In this project we work on the development of an algorithm combining knowledge graphs, automated reasoning over ontologies, and feature learning with neural networks, to generate dense vector representations of biological entities. Through this method, biological data represented as knowledge graphs can be utilized for applications ranging from discovering molecular mechanisms underlying disease over drug repurposing to functional genomics.

  • Predicting protein functions from sequence by a Neuro-Symbolic Deep Learning Model
  • Identifying the function of protein is a key to understand life at the molecular level and has great bio-medical implications. Gene Ontology is a structured vocabulary that captures protein function in a hierarchical manner. However, experimental characterizations of functions is a highly resource-consuming task and cannot scale up to accommodate the vast amount of available sequence data. There are 45000 GO classes structured as DAG resulting into a challenging multi-class multi-label classification problem. The main focus of this work is to develop a neural network based model to annotate a protein function (represented as a sequence of Amino Acid) with GO categories.

  • Machine Learning based Spectrum Prediction in Cognitive Radio Networks
  • Cognitive radio offers the promise of intelligent radios that can learn and adapt to their environment. To date, most cognitive radio research has focused on policy-based radios that are hard-coded with a list of rules on how the radio should behave in certain scenarios. The focus of this work is to formalize application of machine learning algorithms for cognitive radio application and develop a framework within which they can be useful. The work addresses how generic cognitive engine can tackle problems such as dynamic spectrum access and spectrum prediction. Efficient spectrum sensing can be realized by predicting the future idle time of primary users activity in a cognitive radio network. In dynamic spectrum access, based on a reliable prediction scheme, a secondary user chooses a channel with the longest idle time for data transmission. The primary activity in most type of wireless networks governed by several well known network traffic models namely, Poisson, Interrupted Poisson (IP) and Self-similar (SS) traffic. For this work these traffic were simulated and were used in a sliding window methodology with Recurrent Neural Networks to predict future primary activity.