Soochan Lee

Ph.D. candidate in Vision and Learning Lab at Seoul National University

[Soochan Lee @SoochanL33]

Is a sequence model all you need for continual learning? Our NeurIPS 2023 paper "Recasting Continual Learning as Sequence Modeling" points out that continual learning is inherently a sequence modeling problem, and sequence models can be employed as CL solutions. (1/5)

[Soochan Lee @SoochanL33]

@alfcnz Thank you so much for the kind words!!

[Soochan Lee @SoochanL33]

One of my recent works (https://t.co/4dG0i0iQcV) is an example of this direction, and I believe meta-learning can play a crucial role in other fields that require CL, e.g., #ReinforcementLearning. (5/5)

[Soochan Lee @SoochanL33]

Is a sequence model all you need for continual learning? Our NeurIPS 2023 paper "Recasting Continual Learning as Sequence Modeling" points out that continual learning is inherently a sequence modeling problem, and sequence models can be employed as CL solutions. (1/5)

[Soochan Lee @SoochanL33]

#ContinualLearning is NP-hard.

@LauchLab, @syimplectic, and @tommy_da_cat brilliantly proved it in their ICML 2020 paper (https://t.co/fCo6Juk4SQ) using set theory.

This provoked the following thought process that eventually made me depart from conventional CL. 🧵(1/n)

[Soochan Lee @SoochanL33]

This is why I landed on meta-continual learning, which is often described as learning to continually learn. It is a variation of meta-learning that aims to optimize CL algorithms in a data-driven manner, instead of relying on human knowledge. (4/n)

[Soochan Lee @SoochanL33]

@tetraduzione @alfcnz Maybe, but the intention was to utilize the huge poster board area to cover a wide range of audiences. For ~70% of the people, the left half (which is still quite big for a poster) was enough, while the FAQ was for <10% who are willing to dive into the details. 😉

[Louis Kirsch @LouisKirschAI]

Emergent in-context learning with Transformers is exciting! But what is necessary to make neural nets implement general-purpose in-context learning? 2^14 tasks, a large model + memory, and initial memorization to aid generalization.

Full paper https://t.co/yyp9467WgF

🧵👇(1/9)

[David Abel @dabelcs]

Reinforcement learning often studies agents that solve a task, then stop learning. In contrast, continual RL focuses on agents that never stop learning.

But what does "never stop learning" mean? 🤔

Excited to share our #NeurIPS2023 paper on this!

https://t.co/OvU0PjaqJ6

🧵

[Alfredo Canziani @alfcnz]

Best poster award goes to @soochanl33’s «Recasting Continual Learning as Sequence Modeling» 🥳🥳🥳
Nice graphics, clear explanation, appropriate layout. Plz, more of this! ❤️❤️❤️

[Soochan Lee @SoochanL33]

@alfcnz Thank you so much!! 😂

Here is a clean version:

[Soochan Lee @SoochanL33]

For more details, please check our paper at https://t.co/biZNpWv0cX (5/5)

[Soochan Lee @SoochanL33]

Is a sequence model all you need for continual learning? Our NeurIPS 2023 paper "Recasting Continual Learning as Sequence Modeling" points out that continual learning is inherently a sequence modeling problem, and sequence models can be employed as CL solutions. (1/5)

[Soochan Lee @SoochanL33]

By leveraging the meta-continual learning (MCL) framework, we can train the sequence model on multiple CL episodes at the meta-level, i.e., learning to continually learn. In our experiments, we demonstrate that conventional CL < SGD-based MCL < sequence modeling. (4/5)

[Jim Fan @DrJimFan]

Shane Legg, cofounder of DeepMind, explains the importance of adding Search to Neural Networks.

Background: "Move 37" was a pivotal play by AlphaGo during the 2nd of its 5-game series against Go Champion Lee Sedol. AlphaGo placed a stone in a highly unconventional position that even confused experts, who first deemed it an error. Yet as the game unfolded, the move proved to be strategically brilliant, opening up opportunities that only became clear much later.

In a sense, Search increases intelligence at inference. The AI can take more time to deliberate without altering its parameters, and dynamically tradeoff efficiency with deeper thinking.

Video credit: @dwarkesh_sp podcast w/ @ShaneLegg

[Peyman Milanfar @docmilanfar]

It strikes me that AI alignment/safety is like controllability, and AI interpretability is like observability. These are both classical concepts from system theory that seem to be largely unknown to AI researchers.

[Jim Fan @DrJimFan]

Ummm ... why is this a surprise? Transformers are not elixirs. Machine learning 101: gotta cover the test distribution in training!

LLMs work so well because they are trained on (almost) all text distribution of tasks that we care about. That's why data quality is number 1 priority: garbage in, garbage out. Most of LLM efforts these days go into data cleaning & annotation.

[SNUVL @SNUVL]

2 papers accepted to NeurIPS 2023
@NeurIPSConf
🎉Super works from our authors! 📷⬇️