Jiajun He
@JiajunHe614
Followers
101
Following
100
Media
15
Statuses
45
PhD student in machine learning
Cambridge, England
Joined August 2024
[1/9]🚀Excited to share our new work, RNE! A plug-and-play framework for everything about diffusion model density and control: density estimation, inference-time control & scaling, energy regularisation. More details👇. Joint work with @jmhernandez233 @YuanqiD, Francisco Vargas
3
19
71
RT @MolSS_Group: Join us tomorrow afternoon at 4pm (UK time) if you are interested in scalable and simulation-free method for neural SDE tr….
0
4
0
Looking forward to hearing from @GrigoryBartosh Please join us if you are also working on time series data or sequences!.
On the coming Tuesday (July 1st), we will have.@GrigoryBartosh talking about “SDE Matching: Scalable and Simulation-Free Training of Latent Stochastic Differential Equations" ( 🚀, from 4pm to 5pm (UK time). Join us via Zoom 🔥
0
1
5
RT @MoleiTaoMath: Generative modeling data with multiple modalities (e.g.continuous,discrete,manifold,constrained)?. ppl often tokenize eve….
0
31
0
RT @k_neklyudov: Why do we keep sampling from the same distribution the model was trained on?. We rethink this old paradigm by introducing….
0
24
0
RT @majdi_has: (1/n)🚨You can train a model solving DFT for any geometry almost without training data!🚨. Introducing Self-Refining Training….
0
38
0
Really interesting and inspiring work! Congrats.
🧵(1/6) Delighted to share our @icmlconf 2025 spotlight paper: the Feynman-Kac Correctors (FKCs) in Diffusion. Picture this: it’s inference time and we want to generate new samples from our diffusion model. But we don’t want to just copy the training data – we may want to sample
0
0
2
[9/9] ✨RNE also connects and unifies many prior works. Special thanks to the authors of these amazing works! They largely inspire our work! 💡
0
1
5
[8/9] All results and estimators can be seamlessly translated to CTMC-based discrete diffusion models, extending and further showcasing the plug-and-play nature of our framework⚙️!.
1
2
8
[7/9] To improve the stability and convergence of RNE, we introduce a linear Gaussian reference Markov kernel, which removes unstable terms and enhances the accuracy of our results🎯
1
1
4
[6/9] If you are working on EBMs, you can also regularise your energy-based Diffusion with RNE🪢!
1
1
4
[5/9]🤸♀️🤸♂️RNE enables flexible designs for the SMC in inference-time control for various inference-time control tasks, including annealing, reward-tilting, composition, and CFG.
1
1
4
[4/9] Leveraging our RNE, we can compute the density of an SDE with a tractable initial distribution (e.g. Diffusion models). Our density estimators (RNDE and its variants) showcase competitive performance compared to prior work👇
1
1
5
[3/9] Inspired by Bayes’ rule, RNE leverages the ratio between a pair of forward and reverse-time SDEs. We explore.- 🚀Div-free density estimation for SDEs with reversals.- 🤸Unified, derivation-free and flexible SMC!.- 🛠️Div-free energy regularisation for diffusion EBMs
1
1
4
[2/9] arXiv: RNE unifies importance weight calculation & density estimation for diffusion. For example, you can do inference-time control (e.g, anneal/reward-tilt) with flexible denoising process using the same interface👩💻, without redoing maths🤯!
1
1
4
RT @MolSS_Group: Our first session will be on next Tuesday (Jan 3rd) from 3pm to 4pm (UK time) 🚀. This session will be given by @jmhernande….
0
5
0
Join us for more cutting-edge works!.
We’re thrilled to announce the launch of the MolSS Reading Group! 🚀.🔬 MolSS = Machine Learning for Molecular Simulations and Sampling.🔥 First session: June 3rd, 3 PM (UK time) by @jmhernandez233.🗓️ Regular sessions: Every other Tuesday at 4 PM (UK time), 45–60 minutes.
0
0
2
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
Come and see our paper at room Peridot 202-203 at 4:45pm today!.
If you are still at ICLR, consider stopping by our poster at the FPI workshop to visit the poster with my amazing collaborators @JiajunHe614 and Francisco!.
0
0
2
RT @MingtianZhang: Checkout this work on training one step generative model, pre-training is key!.
0
4
0