Shuyu Lin


Researcher.


Traveller.


Food Lover.

About me

I am currently a PhD student at University of Oxford, working to decipher the secret recipe of artificial intelligence

Outside research, I teach, travel and hunt for good restaurants. 

I also care very much about the environment and would love to promote diversity & equality in our society.

Research

header image for WiSE-ALE: posterior distribution of MNIST 2D embedding

WiSE-ALE: Wide Sample Estimator for Approximate Latent Embedding

Shuyu Lin, Ronald Clark, Robert Birke, Niki Trigoni, Stephen Roberts

International Conference on Learning Representations (ICLR) Deep Generative Models Workshop, 2019

arxiv, bibtex

We consider an auto-encoding process w.r.t. the entire dataset, instead of a sample-wise auto-encoding process. We demonstrate that the representation (latent embedding) learnt under such procedure has enhanced reconstruction quality.

image13

Learning Semantically Meaningful Embeddings Using Linear Constraints

Shuyu Lin, Ronald Clark, Robert Birke, Niki Trigoni, Stephen Roberts

International Conference on Computer Vision and Pattern Recognition (CVPR-W) Explainable AI Workshop, 2019

pdf, bibtex

We investigate the use of invertible decoders for learning a semantically meaningful latent space.

Generated samples and associated uncertainty

Balancing Reconstruction Quality and Regularisation in ELBO for VAEs

Shuyu Lin, Stephen Roberts, Niki Trigoni, Ronald Clark

arxiv preprint, 2019

arxiv, bibtex

A principled approach for balancing the reconstruction-regularisation tradeoff in VAE models with a Gaussian decoder. Sensible uncertainty estimate on the generated samples is produced.

Model sketch for consistent auto-encoding

Towards Consistent Variational Auto-Encoding

Yijing Liu*, Shuyu Lin*, Ronald Clark (*equal contribution)

AAAI Conference on Artificial Intelligence Student Abstract, 2020

to be released

We introduce a novel consistency loss that directly requires the encoding of the reconstructed data point to match the encoding of the original data, leading to better representations. 

Food

Home Made Pizza

image20

Mum's Ramen

image21

Matcha Muffin

image22