Iat Chan

University of Oxford

Master of Science (M.Sc.), Deep Learning, Computational Linguistics, Bayesian Statistics

January 1, 2014 – January 1, 2015

Tsinghua University

Bachelor of Engineering (BEng), Computer Science, High Performance Computing

January 1, 2010 – January 1, 2014

Qube Research & Technologies

Quant Researcher

May 1, 2026 – Present

イギリス ロンドン エリア · On-site

Stealth Startup

Quantitative Researcher

October 1, 2022 – April 1, 2026

OakNorth

Tech Lead of Machine Learning Team in OakNorth

January 1, 2020 – October 1, 2022

OakNorth

Machine Learning Engineer

June 1, 2019 – January 1, 2020

Bloomberg LP

Research Scientist - Machine Learning

November 1, 2015 – June 1, 2019

London

University of Oxford

MSc in Computer Science

August 1, 2014 – September 1, 2015

Tsinghua University

BEng in Computer Science & Technology

August 1, 2010 – July 1, 2014

th4j - A wrapper of Torch TH library for Java (and other JVM langauges).

September 1, 2015 – November 1, 2015

This project aims to construct a numerical computing library, based on tensor (i.e. multi dimensional array), for benefiting scientific computation on JVM platform. This project stands out comparing with other companions as (1) Based on TH (and THC), which is constructed on OpenBLAS (or Intel MKL), it enables us to play with the tensors in torch, a recently popular deep learning toolbox, directly in JVM. (2) Multi dimensionality support, other frameworks such as BIDMat or Breeze focus on 2-d (maxtrix) or 1-d (vector) tensor. While most practical neural network model require high-dimensional tensors (e.g. ConvNet for Pictures needs 4-d tensor). (3) Fast computation on GPU, the implementation of most other frameworks (e.g. BIDMat, ND4j) consists of many explicit synchronizations, while operations on THC (the cuda library for torch), are mostly asynchronous, maximizing the performance. (4) Written in Scala, the manipulation of tensors can be expressed in very concise and clean manner.

Input Method Engine by LSTM Recurrent Neural Network

February 1, 2015 – August 1, 2015

Experienced PC users do not need to stare at the keyboard to type, as touch-sense feedback of real keyboard makes people know how to coordinate the fingers to find the right keys to tap. However, in the context of virtual keyboard, a major input method for small electronic devices (e.g. Smart Phone, Tablet), even if you are a typing master on real keyboard, unless you stare at the virtual keyboard constantly, you will easily enter a sequence of gibberish while typing. Key observation here is that without touch sense (e.g. F,J on real keyboard) and visual feedback, people will fail to `sense' the correct tapping locations, leading to significantly high error rate on virtual keyboard. Yet, the tapping locations on screen still have patterns. In this project a bayesian network is constructed based on empirical study in the field of human computer interaction to model tapping patterns. As exact inference is intractable, due to each tap conditioning on all previous taps, and sampling based approach is too slow, a long short term memory recurrent neural network is used as variational family and variational inference technique is employed.

Automatic song/speech segementation

July 1, 2013 – August 1, 2013

This system aims to automatically mark which parts of audio clips are speech while which parts of them are song.

Fantastic Drummer

September 1, 2012 – January 1, 2013

This project aims for automatically generating notation for songs without human interference, so that the processed songs are playable in famous rhythm games such as taiko no tatsujin(太鼓の達人).