Care to elaborate how AZ/MZ are not model based? MuZero learns a transition function. AZ uses a known transition model to do planning, but I would still say that's model-based RL.
Data-driven robot control methods for solving furniture assembly.
It's an interesting problem, requiring both dexterous manipulation and long-term planning. It's also compositional, so I believe some form of hierarchical control and planning can solve it.
Hey all, I'm one of the authors of the environment. I'm excited to see what researchers will come up with to tackle the furniture assembly problem. Let me know if you have any questions or comments.
We don't plan too many meetings. Once a week for a lab wide meeting. Most of these meetings are casual and just discussions between cliques of phd students on specific topics.
Being in the lab helps me focus on my thinking. I also discuss my current thinking and problems with my peers for critique, and I can do the same for them. I'm more conditioned to do research, and my mind wanders less in the lab.
However, I do agree that having free time and being away from the lab is important to get a fresh view of things.
Disclaimer: Our field is extremely competitive (deep learning) and other labs may have better working styles depending on their lab size and member composition.
My lab's workload on average is 8 - 12 hours a day for 6-7 days a week, ramping up even more closer to deadlines. That being said, a lot of this time (50% or more) is for thinking and discussing. This schedule, while hard, seems to be effective for publishing in good conferences.
Good question, for me I programatically generate a folder with the hyperparameter key - values in the name and store the checkpoints under it. As you can imagine, it can get out of control quickly if not managed well, but it works for 1-3 person projects. For anything more large scale or organized, I would recommend looking into Comet ML which lets you query and filter your experiments by hyperparameter ranges instead of manually looking at folder names.
seems great for my reinforcement learning models. instead of parsing my hyperparameters through the tensorflow cli API and editing the training file a line at a time to take in an additional hyperparameter, I can just directly set them through the cli with fire.
Depends on how simple you want to get. If literally displaying text once in a while, you don't even need jekyll. If you are a blog writer, want some sort of templating / reusable snippets, or want cool themes then jekyll is good.
I see, but I think for areas like Reinforcement Learning or sequence models though eGPUs would be significantly slower since you have to constantly shuttle new data into the GPU from your computer.
Total compensation: Definitely a wide range, but a few hundred K would be the median.
Recruiting: one of my labmates flat out refused to do algorithmic interviews for a research position and the company still gave him an offer. Your projects, papers, and ability to demonstrate your understanding of research topics are far more important then your ability to solve a small, defined problem in 45 minutes.