Jihao Andreas Lin
@JihaoAndreasLin
Followers
105
Following
51
Media
4
Statuses
23
PhD student in probabilistic machine learning at University of Cambridge
Cambridge, UK
Joined October 2022
"Improving Linear System Solvers for Hyperparameter Optimisation in Iterative Gaussian Processes". Three techniques to accelerate marginal likelihood training in GPs by up to 72x without sacrificing performance!. Check out our paper here: . (1/6).
2
6
51
RT @jmhernandez233: Looking for Postdoc/Research Assistant to work on deep generative models, with a focus on the domain of molecules. App….
0
22
0
RT @TonyRKOuYang: New revision of “BNEM: A Boltzmann Sampler Based on Bootstrapped Noised Energy Matching” 🚀. (1/6) We introduce NEM and BN….
0
10
0
RT @runame_: 1/7 Still using Adam?. If anyone wants to try a distributed PyTorch implementation of SOAP/eigenvalue-corrected Shampoo with s….
0
85
0
RT @kayembruno: Diffusion models are so ubiquitous, but it's difficult to find an introduction that is concise, simple and comprehensive.….
0
116
0
RT @yichao_liang: How can we get VLMs to help robots solve complex long-horizon tasks? Introducing VisualPredicator: an agent that leverage….
0
17
0
Many thanks to my collaborators @shreyaspadhy, @kayembruno, @JaviAC7, and @jmhernandez233 for this amazing project!. Check out our paper here: (6/6).
0
0
6
With large datasets, solvers require too much compute to converge. Instead, they are stopped after a limited compute budget. Combining early stopping and warm starting, solvers accumulate progress across marginal likelihood steps, amortising the costs of linear solves! . (5/6)
1
0
4
Linear system solvers are typically initialised at zero. However, each marginal likelihood step only changes the kernel hyperparameters and linear systems slightly. Therefore, we can reuse the solution from the previous step as initialisation to warm start our solvers! . (4/6)
1
0
5
With the standard gradient estimator, solvers require more iterations to converge as the model fits the data during optimisation. Using the pathwise gradient estimator, the required number of iterations is constant. You also get posterior predictions as a by-product! . (3/6)
1
0
4
Iterative GPs use linear system solvers, like conjugate gradients, alternating projections, or stochastic gradient descent, to construct estimates of the marginal likelihood gradient. However, solvers dominate the computational costs! How can we make them faster?. (2/6)
1
0
2
RT @CambridgeMLG: "Stochastic Gradient Descent for Gaussian Processes Done Right" .🎓 @JihaoAndreasLin*, @shreyaspadhy *, @JaviAC7 *, @austi….
0
3
0
RT @shreyaspadhy: If you're interested in sampling from Gaussian Processes on millions of datapoints in linear time using SGD, check out ou….
0
13
0
RT @runame_: How can Kronecker-Factored Approximate Curvature (K-FAC) be generalised to modern deep learning architectures like transformer….
0
21
0
RT @kayembruno: We'll be presenting our spotlight poster on accelerating molecular dynamics simulation with deep generative models at #Neur….
0
9
0
RT @CambridgeMLG: We're at #NeurIPS2023 and excited to share our work! Our group and collaborators will present 22 papers at the main confe….
0
20
0
RT @CambridgeMLG: Looking to do a PhD in machine learning? Want a collaborative research environment with world-renowned researchers? 🧑🎓 W….
0
29
0
RT @runame_: My supervisor Rich Turner has written an introduction to transformers. I highly recommend anyone wanting to learn (more) about….
0
188
0
RT @CambridgeMLG: We're at #ICML2023 and excited to share our work! Our group and collaborators will be presenting 8 papers at the main con….
0
7
0