Thanks.

Wow! Also thanks for being so open. Good luck :)

Good grief , Get to the point !

cheeky video title

Key Features (Or Maybe Not) of Open-Endedness: Innovation: The system continually generates new, innovative solutions rather than converging on a single "best" one. Diversity: Over time, the system produces a diverse set of solutions that can perform various tasks or adapt to different environments. Complexity: The complexity of the evolved entities often increases over time, mirroring the complexity seen in biological evolution. **Dog**: Hey Whiskers, what's got you so wrapped up today? More string theory? **Cat**: Ah, Rover, always with the puns. No, no, I've been pondering about this thing humans call "Open-Endedness" in algorithms. **Dog**: Algorithms? That's way above my pay grade. I'm just here for the treats and belly rubs. **Cat**: Classic Rover. But listen, this is like the ultimate game of fetch, except it never ends and the stick keeps morphing into new things! **Dog**: Morphing sticks? Now you're talking my language! Go on... **Cat**: Imagine you fetch a stick, but then it turns into a squeaky toy. And then maybe into a Frisbee! **Dog**: Whoa! A never-ending game with ever-changing toys? That's like, a dream come true! **Cat**: Exactly! It's like us felines having a laser pointer that turns into a mouse, a butterfly, and then maybe a little robot we can chase around! **Dog**: I can't even imagine a fetch game that cool. What's the catch? **Cat**: There's no catch. That's the whole point! The algorithm keeps creating new things to chase, or in human terms, new solutions to problems they didn't even know they had. **Dog**: So it's like digging a hole and instead of finding dirt, you find bones, toys, and maybe even a treasure chest? **Cat**: You're catching on, Rover! It's all about endless possibilities. **Dog**: Mind blown, Whiskers, mind blown. Let's pitch this to the humans; maybe they'll build us a toy like that! **Cat**: Ha, as if we could make them do anything! But hey, a feline can dream, right? **Dog**: And a dog can always hope for a better game of fetch! To infinity and beyond! **Cat**: Alright, Buzz Lightyear, calm down. Let's just stick to dreaming about endless games and let the humans figure out the rest. **Dog**: Deal, Whiskers. But if this thing ever becomes real, I get first dibs on the morphing stick! **Cat**: Only if I get the first swipe at the shape-shifting laser pointer! **Dog**: You've got yourself a deal! **Cat**: Pawsome! Now, back to my existential pondering. **Dog**: And I'll go back to digging holes, maybe I'll find that treasure chest after all!

hard to follow what is being said and connect/ground it on the slides! better presentation skills may help.

All these experiments are in a grid-world setting. Would these hold up when when we are dealing with an Robot acting in real world ?

Amazing talk!

Audio issues

At 51:36 there's a cut in the video. Is this intentional?

About the question whether PAIRED is doing more than Domain Randomization: If you get a policy that adapts to all suggested environments proposed by DR, it might still not be able to generalize to environments outside of the domain of what the DR is capable of right? Because it could have memorized all the proposed environments. But with PAIRED we constrain the situations the agent would encounter and in that sense force it to learn skills that (hopefully) do generalize better?

Very interesting, and I even think that some of the issues presented about answering counterfactual queries could also be problematic for what was described as "constraint policy" methods in general

Под Херсоном разбомбили целую колону техники ,куча трупов с обеих сторон , остановитесь!!!

Thank you for this enriching presentation

🔥

Hi Sergey, Thanks for the nice lecture but It cab be improved by explaining how to create our own custom environment for CQL using offline data because I thought the real challenge is here to design the ENV. It will be great if you can share some lecture or link on that.

Cross attention for whole part relationships of solid objects will be a sure shot way of easiest algorithm for CV.

Bounding boxes is how i guess Tesla is doing CV, its the easiest way to go about.

Special cross attention slotting can be infinite. To make them finite. Their 2 d representations in basic geometric shapes can be much easier.. Like if it a square or circle or triangle, with their 3d feature of height. Based on this a computer can work to find the way based on laws of motion of these objects which are luckily quite few. Also, energy functions of Yann can be used to check the motion of objects, Usually those who move more will have higher energy changes or in my terminology exchanges

Great talk and an inspiring idea from Sam.

Any recommendation for how to apply reinforcement learning on continuous space(not discrete) data.

Where's the code? #papersandtalkswithcode

Speaking nicely

Speaking nicely