Asif Khan

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

I am a third-year PhD student at the School of Informatics, University of Edinburgh. I am a member of the Bayesian and Neural Systems group, advised by Prof. Amos Storkey and Prof. Chris Williams. I am broadly interested in Unsupervised Representation Learning, Probabilistic Models, Bayesian Inference, and Physics-prior in Neural Networks. My current research is on using Hamiltonian operators to disentangle motion from content in sequential data. Previously, I completed MSc. in Computer Science at the University of Bonn, where I was also a research assistant at the Smart Data Analytics (SDA) group. I worked on Machine Learning for Knowledge Graphs with Prof. Asja Fischer and Prof. Jens Lehmann. In my master thesis, I developed a GAN framework for speech-to-speech synthesis advised by Prof. Asja Fischer and Dr. Fabien Cardinaux.


Teaching

University of Edinburgh

Tutor and Marker for Probabilistic Modelling and Reasoning 19/20
Marker for Machine Learning Practical 19/20, 20/21
Marker for Introductory Applied Machine Learning 19/20, 20/21
Marker for Data Mining and Exploration 20/21

University of Bonn

Teaching Assistant for MSc. course on Knowledge Graph Analysis, WS17/18, WS18/19
I was responsible for preparing and delivering tutorials and marking exams. I developed programming and theoretical exercise for the course.Link.

LNM Institute of Information Technology

Teaching assistant for C Programming Lab, Fall13
Teaching assistant for Digital Signal Processing Lab, Fall15

Projects

  • Soccer Ball Localization
  • This work addresses the soccer ball localisation problem using deep convolutional neural networks (CNN). We use fully convolutional networks (FCN) to predict the probability map of a ball's location from an image. Later, we extend it to the video data using temporal models, namely ConvLSTM and TCN, on top of the base FCN model. Instead of training from scratch, we take advantage of transfer learning and finetune FCN. We evaluate all our experiments on a novel dataset prepared as a part of this work. Furthermore, we present empirical results to support the effectiveness of using the ball's history in challenging scenarios.

  • Game Artificial Intelligence
  • In this project, we address three AI problems associated with game design: a) strategies for turn-based games, we propose a probabilistic scheme based on the winning statistics of placing a move in a particular cell, b) game trees & path planning, we analyse the complexity of game trees in terms of branching factor, leaf nodes and non-terminal nodes. We show the computational cost of using minimax to find an optimal move in a high branching factor game like Connect4; we then look at an alternative version of minimax known as depth restricted search. In path planning, we study the cost of uninformed and informed search methods on several 2D game maps. c) behaviour programming, we developed a fuzzy logic controller for the breakout. We further proposed an algorithm using self-organising maps and Bayesian imitation learning for predicting player trajectory in a Quake III game map.

  • Commonsense Knowledge Base Reasoning
  • Common sense knowledge can improve the performance of several tasks like text classification, sentiment analysis, entity resolution, question answering etc. ConceptNet is a database representing the various form of commonsense knowledge as a multi-relational graph. In this work, we use a relational learning approach to predict the missing links in a database and thus increase the coverage of ConceptNet. We do so using tensor factorisation and neural networks methods (RESCAL, DistMult, ERMLP). We further combine the extended ConceptNet database with natural language sentences to learn common sense enriched word embeddings.

  • Multi-Modal Learning for Disease-Gene Prioritization
  • Identifying gene-disease associations plays a vital role in better understanding the gene functions, interactions, and pathways, thus critical for improving medical care. In this work, we develop a novel method for learning embeddings of disease and gene entities from text data and biological knowledge graphs. We further use the embeddings to discover novel gene-disease associations.

  • Learning Representation of Biological Knowledge Graph
  • The main success of the Semantic Web in the life sciences is due to 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 validating and inferring new biological relations or analysing 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 develop an algorithm combining knowledge graphs, automated reasoning over ontologies, and feature learning with neural networks, to generate vector representations of biological entities. Through this method, we can utilise the biological data represented as knowledge graphs for applications ranging from discovering molecular mechanisms of 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 (CR) offers the promise of intelligent radios that can learn and adapt to their environment. To date, most CR research has focused on policy-based radios that use hard-coded set of rules to decide how the radio should behave in specific scenarios. In this project aim is to formalise the use of machine learning algorithms for CR applications and develop a framework within which they can be useful. Specifically, our work addresses how a generic cognitive engine can tackle Dynamic Spectrum Access (DSA) and Spectrum Prediction (SP) problems. We develop a model for efficient spectrum sensing by predicting the future idle time of primary users activity. Based on a reliable prediction scheme in DSA, a secondary user chooses a channel with the longest idle time for data transmission. Several well-known network traffic models: Poisson, Interrupted Poisson (IP), and Self-Similar (SS), govern the wireless networks' primary activity. We simulated these traffics and used Recurrent Neural Networks for predicting future primary activity. Later, we collected our approach on real traffic data.