Thang Doan

18 posts

Thang Doan banner
Thang Doan

Thang Doan

@doan_tl

Pr. Applied Scientist @Oracle | Postdoc @Mila_Quebec, PhD @mcgillu, MSc @ucl

LAX Katılım Aralık 2020
153 Takip Edilen135 Takipçiler
Thang Doan retweetledi
Roberta Raileanu
Roberta Raileanu@robertarail·
CSP accepted as spotlight at #ICLR2023. Kudos to @jb_gaya who led this project and to all the other wonderful co-authors @doan_tl @LucasPCaccia @LaureSoulier @LudovicDenoyer
Roberta Raileanu@robertarail

CSP: our new algorithm for scalable continual reinforcement learning. CSP allows users to control the trade-off between performance and model size by adaptively learning a subspace of policies. SOTA on Continual World and Brax. Interactive Website: bit.ly/3tRb9ri

English
4
8
47
11.3K
Thang Doan retweetledi
Roberta Raileanu
Roberta Raileanu@robertarail·
Building a Subspace of Policies for Continual Reinforcement Learning When: Fri Dec 9, 8:25am PST Where: Virtual Paper: arxiv.org/abs/2211.10445 Website: bit.ly/3tRb9ri Thread: twitter.com/robertarail/st… led @jb_gaya w @doan_tl @LucasPCaccia @LaureSoulier @LudovicDenoyer
Roberta Raileanu@robertarail

CSP: our new algorithm for scalable continual reinforcement learning. CSP allows users to control the trade-off between performance and model size by adaptively learning a subspace of policies. SOTA on Continual World and Brax. Interactive Website: bit.ly/3tRb9ri

English
1
4
11
0
Thang Doan
Thang Doan@doan_tl·
We propose a variant of Orthogonal Gradient Descent (OGD) which leverages the structure of the data through Principal Component Analysis called PCA-OGD. Our method is twice as memory efficient as OGD. (4/n)
Thang Doan tweet media
English
1
0
1
0
Thang Doan
Thang Doan@doan_tl·
We introduce the “NTK Overlap matrix” as a task similarity measure that governs the CF. Unlike Vanilla SGD, projection based methods mitigate CF by operating orthogonally to previous task subspaces. This reduces the overlap between tasks leading to less forgetting. (3/n) 👇
Thang Doan tweet media
English
1
0
0
0
Thang Doan retweetledi
Ankesh Anand
Ankesh Anand@ankesh_anand·
The RL formalism is powerful in its generality, but poses a hard problem: how can we design agents that learn efficiently & generalize well, given only sensory info and a reward signal? Self-supervision might be the answer, join us at the ICLR workshop: sslrlworkshop.github.io
Ankesh Anand tweet media
English
3
61
313
0
Thang Doan retweetledi
Sébastien Couasnon 
Sébastien Couasnon @SebCouasnon·
Population déjà vaccinée : 🇮🇱24% 🇬🇧5,5% 🇺🇸4% 🇩🇰2,5% 🇮🇹1,7% 🇪🇸1,6% 🇨🇦1,3% 🇩🇪1,2% 🇪🇺1,1% 🇵🇹1% 🇷🇺1% 🇫🇷0,6%
Français
14
28
109
0
Thang Doan retweetledi
Vincenzo Lomonaco
Vincenzo Lomonaco@v_lomonaco·
🔥 The @ContinualAI Reading Groups are back! 🔥 This Friday, 17.30 CET we will host Mehdi Abbana Bennani (Aqemia) who will present the paper: "Generalisation Guarantees for Continual Learning with Orthogonal Gradient Descent" eventbrite.it/e/generalisati…
English
1
1
7
0

Keşfet

@LucasPCaccia @LudovicDenoyer @LaureSoulier @robertarail @jb_gaya @ContinualAI @v_lomonaco @bogdan_mazoure