B (2:05)

Yeah. And I don't know if any of you guys want to want to chime in on other elements of coming into like deciding on this problem.

D (2:14)

So it's like broadly my lab works on deep reinforcement learning, but historically deep meant like two or three or four layers, not 1,000. When Kevin Yishon mentioned they wanted to try really deep networks, kind of skeptical it was going to work. I've tried this before. It doesn't work. Other papers have tried this before and it doesn't going to work.

C (2:34)

So I was very, very skeptical starting out.

D (2:36)

I don't know if I conveyed this at the time, but that was my prior going in because.

B (2:41)

But do you, do you view your job as like screening or like, hey guys, this is probably isn't going to work. You should try a different idea, you know, like, or should you be encouraging even if it's dumb, it's selecting bets. Yeah.

D (2:52)

And this was a bet I was willing to make.

B (2:55)

What made you willing to make a bet?

D (2:57)

It seemed relatively low cost in that Michal in particular had spent the past year developing infrastructure that made it a lot easier to run some of these experiments. And the precedent was deeper. Now we should do a whole lot better. That's what the deep learning resolution revolution has been over the last 10 years.

B (3:15)

Yeah, I know. Why do we stop making them deeper?

D (3:17)

And reinforcement learning was like this one anomaly where we continue to use these really shallow networks. And that's particularly true in the settings that we were looking at, where you're starting from scratch, you're starting from nothing.

B (3:27)

Any other perspectives you guys want to chime in with?

C (3:30)

I guess maybe I should just go over like an overview of our project.

B (3:33)

Yes, okay, sorry. Yes.

C (3:35)

So the way that I kind of view our project is that if you look at the landscape of deep learning, you know, you have nlp, like language vision and then rl. And as Ben kind of alluded to, you know, like in, in language, in vision, we've sort of converged to these like paradigms of scaling to massive networks. Right? Like hundreds of billions of parameters, trillions of parameters. And there's been, you know, a lot, a lot gained in deep learning from, from that. Right. And then, but then it seems like in the third sort of branch of deep learning in deep rl, that has not yet been the case. Like I was very surprised, like coming into some like, you know, Ben's class and seminar when I was looking at the networks. Oh, why were you just using like a simple like 2 layer MLP for like these frontier sort of, you know, state of the art RL algorithms. And so I was very curious, like, can we design RL algorithms? Can we sort of put together a recipe for RL that can allow it to scale in potentially, you know, analogous ways that language envision might scale. And so what we did is that we know that traditional rl, like, let's say like value, value based RL doesn't really scale. Right. This is pretty clear from the literature. So we tried a different approach at RL called self supervised RL where instead of learning a value function, we're learning representations of states, actions and future states such that the representations along the same trajectory are pushed together, the representations along different trajectories are pushed apart. And this is just a different approach to RL that allows us to learn in a self supervised manner so we can solve task reach goals without any human crafted reward signal. And so we know that self supervised learning is scalable in these different areas of deep learning. So can self supervised RL scale in similar ways? When we first tried it, it actually didn't work. We made the network steeper, the performance totally degraded. But then separately was like there's also some other work in our literature. We tried residual connections and then there's a few other architectural components that we had to put into the recipe. And then all of a sudden one day I ran this experiment and there was this one environment in which there was going from doubling the depth, didn't really do anything, but doubling the depth again with these different components suddenly skyrocketed. Performance in this one environment, getting this.

D (5:45)

To work was very non trivial in the sense that usually when we think about doing hyperparameter optimization, we try changing A, see if it makes it better, try cranking B, see whether it makes it better. And, and if we just made the depth bigger, it makes it worse. We just had residual connections, didn't make it better. And it was really this combination of factors that Kevin and Ishan figured out that really made this work.

E (6:07)

And as a precursor to that, we also tried scaling along different dimensions. So scaling the batch size, scaling the width of the networks or the hidden layers. Yeah, pretty much kind of similar to just scaling depth naively. And then once we started introducing residual connections, layer norm, these specific architectural choices, that's when we saw significant jumps in performance like these critical depths at which performance multiplies by a pretty huge factor. And that's where we really notice unlocking some significant performance gains as opposed to scaling just along with which did yield some performance improvements. But when you look at the number of parameters that your network has as you grow width, it's roughly a quadratic as opposed to something like growing depth. So it's more in some senses more parameter efficient, also more sample efficient from the experiments that we conducted.

B (6:54)

Nice. In some ways you're sort of replicating stuff that is seen in the wild, but on a very small model that you can study, would you say that's.

E (7:04)

Yeah. So kind of to what Kevin said earlier, we saw these huge performance improvements in language models, image generation models, by making them larger, making them deeper, which seems very intuitive. And so that's why our work, we draw from foundational research like residual networks, which employ residual connections to avoid vanishing gradients. And that's something that we show in some of our ablations in our paper, further down, probably in appendices where he did experiments without these residual connections. And so sort of borrowing these concepts that have existed in other fields and applying them to this setting with RL and showing that it works.

B (7:40)

Before Ben has to go, I'll leave the sort of last word to him. What additional work does this inspire that you want to push on next?

D (7:49)

I think there's one thing I'd clarify about the paper and then I'll directly answer the question. Yes, I think the thing I might clarify about the paper is I think a lot of people reading the title are like, wow, big networks, they're great. I'll take big networks.

B (7:59)

You solved it. Now we can just go, yeah, we.

D (8:01)

Just take big networks, add them to ppo, add them to sac, add them to your favorite reinforcement learning algorithm. But I think that's actually not the main conclusion. I think the main conclusion is that using big networks not only requires these architectural tricks, but also, as Kevin mentioned before, it requires using a different objective. This objective doesn't actually use rewards in it. And so there's another word in the title, reinforcement learning, that also might be a little bit of a misnomer because we aren't directly trying to maximize rewards. Our code doesn't have a line of code saying maximize rewards here. And so is at the end of the day, this a reinforcement learning method? I don't know. It looks much more similar to the self supervised methods and in other areas of machine learning. And so I think that the method and the work really stands in some sort of interesting intersection of reinforcement learning and self supervised learning research. And we had this little figure on the bottom left of the poster, which was a screenshot of a slide from Yann Lecun talking about how to build intelligent systems and whether that's going to be done by unsupervised learning or supervised learning and reinforcement learning. And I think what her paper really suggests is that the boundary between these things is really blurry and maybe the key to building intelligent systems are going to be leveraging insights from all of them.

B (9:16)

Yeah, the layer kick.

D (9:17)

Exactly.

B (9:19)

Well, thank you for your time. I Know you have to go soon, unfortunately, after Ron. Yeah, thank you so much for coming. I think that insight of blurring things is interesting. I don't know if you were talking about the abstraction layer of representation learning. I don't know if that triggers anything in terms of the mix between self supervised and reinforcement learning. Is that something fundamental that you've discovered or that people don't understand when they read the paper?

C (9:45)

Yeah, I think the best way that I would explain it is that we know that standard RL is not super scalable. And so why can this different approach or different objective RL be scalable? I think it's because we're fundamentally shifting the burden of learning from something like Q learning or regressing to TD errors, which we know is quite spurious and noisy and biased, to fundamentally a classification problem. We're trying to classify whether future state is along the same trajectory or along a different trajectory. And we do this with representation learning. And we know that classification cross entropy loss and representation learning is scalable in the deep learning literature, if we think about language and, and some of the objectives there. So in some sense we're kind of blurring the lines. We're doing reinforcement learning. It's still an actor critic reinforcement learning algorithm. It's like a goal condition reinforcement algorithm, but the objective, the burden of learning, of solving that RL task shifts to something that's more similar to objectives that you might see in language and vision that we know have scaled so much. And so I think that's one of the fundamental insights that we've seen is that it seems like by approaching RL in this different approach, we're able to get so much more out of. We were able to scale our networks significantly beyond what is standard used in rl.

A (11:02)

Can I jump in? I will just give a bit more of context about the architecture because we use another objective, the influence or the contrastive loss. However, the architecture is quite similar to the previous works of previous papers like Bro or Simba Simba Foul and Simba 2. Simba 4 1. Simba 4 2. So we also tweaked a bit of this architecture. However, it's not that we invented the wheel for the first time, it's the merging between the architecture and the objective that makes the scale really go up and performance follow the scale.

B (11:46)

I think that's something that we should probably mine deeper. Do you think? I guess what domains, what industry, you've applied it on multiple different types of networks or data sets. Is there a particular affinity that you think is kind of low hanging food.

C (12:03)

Yeah. So actually if you look at a lot of our tasks, they're particularly sort of like robotics tasks. So this is. Personally I'd be very curious about how a work like this could impact the robotics field. My understanding of robotics is that a lot of robotics are. Now there's kind of a few different approaches. One approach is we want to train robots using imitation learning. So we try to collect an insane amount of data, we have a ton of Cuban supervision and we try to scale up this data and we're learning with imitation learning. But on the other hand, perhaps there's another approach which is, for example, goal condition reinforcement learning, where we can actually train robotic agents, controlling RL agents to solve meaningful tasks with absolutely no human supervision. No demonstration.

B (12:45)

It's much more scalable.

C (12:46)

Yeah. So yeah, so this could serve as an alternate approach and perhaps instead of like scaling data, like scaling manual human supervision, which is, you know, not super scalable, if there are ways to sort of make goal condition reinforcement learning scalable. And like we can just scale the architecture or we can scale.

B (13:02)

Because you're focused on other objectives.

C (13:03)

Yeah, right. With certain different objectives, I think that could be very exciting. And see how to see how that can affect a field like robotics, for example.

B (13:09)

Yeah. Double click on just one thing on the efficiency which Igor is talking about, I would expect the deeper it is, it should be quadratically worse. I'm not familiar with the pre existing literature, I'm just sort of working your intuitions. But basically what are the trade offs that you've found that I think you might want to warn people about? Because you are the guy who mentioned efficiency.

E (13:35)

Sure. Yeah. So I was referring to one of the figures on our poster also in our paper, where we compare the number of parameters that models have as we scale along the axis of depth and as we scale along the axis of width from our baseline architecture, the most baseline one would be a width of 256. The hidden layers have 256 neurons, and then the depth is for four layers or hidden layers. And so the point I was making there is that when you scale along depth, the number of parameters that your model has is going to grow roughly linearly. Whereas with width you're making your network outputs wider. And then the input to the next network is also growing as well. And so the number of parameters your network is then going to have grows approximately quadratically. And so one of the experiments we did was sort of examining as we grow the number of parameters in our model by scaling along These two different choices, which one for the same approximate number of parameters yields a better performance. And the depth curve kind of goes like this, it jumps up pretty fast. That's present throughout our paper for with it grows a little bit more slowly. And so the kind of takeaway from that is that if you are a bit more resource constrained, scaling along depth might be better because with fewer parameters, with a smaller model, a smaller number total learnable parameters.

B (14:46)

Width is expensive.

E (14:47)

Width is expensive. Exactly. And in general, of course, like more parameters is also going to be more expensive. So that's just like another consideration to think about when using these networks, I suppose.

B (14:56)

Yeah. Any other sort of rules of thumbs like that that I can extract that this is just the most basic one that I could think of. I don't know if there's any others.

C (15:05)

Yeah, I guess to your original question of the trade offs, one of the limitations that we say is obviously if you make the networks bigger, it will take longer to run. Right. So if you double the depth at some level of depth, it might take like twice as much to make a forward pass through the network. Right. However, this is not so within our paper. For most environments we are able to saturate, get to almost perfect performance within just. We don't even need to get to a thousand layers. Maybe just 64 layers, for example, is sufficient. And in this regime, the latency of the network is not necessarily, actually even not necessarily a significant bottleneck. Like you can imagine. There's a lot of tasks in which, especially in rl, that collecting data might be the bottleneck. Right. And making four passes through our network may not be the bottleneck. And so in our environment, in our research we specifically used the JAX GCRL environment, which is a JAX based GPU accelerate environment. So we can collect thousands of environment trajectories in parallel at the same time so that we're able to make.

B (16:10)

Oh, this is built in, right?

C (16:12)

This is built in so that we can collect a thousand trajectories at the same time along all these environments and so make sure that we have enough data to saturate the learning from.

B (16:23)

Wow.

E (16:23)

Yeah, that's like work David columns.

B (16:26)

Okay. And I don't know if you want to expound upon that on the jax. Maybe you know, most people are familiar with Pytorch, maybe less familiar with jax.

A (16:35)

With jax, I think JAX is getting the, the traction, especially in RL field because for online reinforcement learning, getting as much data as you can is the most important.

B (16:47)

There's got to be a PYTORCH equivalence. But anyway, any tips for other people also exploring this kind of rollout?

A (16:54)

Yeah, yeah. So I think I can also recommend like for low conditioned rl I'm recommending JAX rl, but there are also like multi agent JAX implementations and others. So going back to our paper, if you look at the PLOSKI on only see this huge performance increase when we cross like 50 millions of transitions gap. So I think the data is crucial.

C (17:21)

Here, I guess even to build on that. I like drawing analogies to successes in other areas of deep learning, for example in large language models. The reason why we're able to scale to such large networks is that we found a paradigm in which we can leverage the entire Internet scale of data alert. Right. And so data in RL traditionally has been hard to come by, but now with these like GPU accelerated environments, we can collect hundreds of millions of time steps of data within just a few hours. And so I think that this serves as like a really good test bed for us to be able to also find ways to scale up like network capacity and get similar kind of gains.

B (17:57)

I think I asked, are you saying that you have a difference? You would do pre training differently in LLMs. Like what's the, what's the difference objective now? Yeah, very simply the paradigm that you're referencing is next word or next token, right? It's very robust. How do you change that?

C (18:19)

Oh, I'm not saying that I want to leverage insights from that to apply to rl.

B (18:25)

I feel like you should go the other way.

C (18:26)

You think you should go the other way? Maybe. I mean that would be a very interesting visualization to. But actually even on that point, one of the things I was thinking about is that the way that RL objective works is in some sense it's not exactly next word prediction, but it's kind of like next state prediction. You imagine you're at some current state and you're at a current action. And we want to predict whether or not this future state, this certain state is a future state along the same trajectory or a different trajectory. And so in some sense we are actually doing some sort of implicit world model. Implicit, you know, like I don't know.

B (18:59)

If that's a bad word or like.

C (19:01)

In language you do a cross entry loss to classify the next token. Right. And here we're just doing a binary classification of like whether or not some next state is some classification.

E (19:09)

Yeah, yeah.

C (19:10)

And so I do see that there are some like sort of parallels here that perhaps we should dig into deeper and see like what is the core to of what enables deep learning to scale? And then how can we like leverage that? How can we distill those like insights and then apply those across like all different fields, whether it's language or reinforcement learning. Yeah.

B (19:30)

Did you get my meaning about the world model stuff?

E (19:32)

Yeah, yeah. Actually I think I might have heard Professor Eisenbach yesterday talking about this at a poster and he's explaining to a couple of people that because this is doing representation learning and trying to learn these meaningful representations for a given state in action, but for a given goal, in some sense you can think of it almost like learning a model of the environment, learning a model of the world, but without having to do any sort of like next frame prediction or stuff like that. That's a little bit more high dimensional and complex.

B (19:58)

Yeah, I would think the angle that I'm trying to think about and push is instead of learn the next world, they're basically generate a number of candidates, possible worlds and classify them to your point, which is exactly how I do things. Let's say I'm playing poker and I'm trying to classify what hands you have. Well, there's a range of hands based on what you're doing. And the more information I get, the more I resolve to oh, I know exactly what happen you have based on what you're showing or you're bluffing. But that's a different thing. But you know what I mean. I feel like that is the ultimate sort of end goal of representation, which is a world. But I don't know if that is too vague compared to the more concrete types of world models that let's say the video gen people are doing. And then I guess one other thing I'm also exploring. You mentioned the deep models being slower or more expensive. Yeah, that is a trend in the inference world of making models shallower.

E (20:56)

Right.

B (20:56)

And I wonder if this short catchphrase I was thinking about deep teacher, shallow student would be a good deployment paradigm. You push the frontier capabilities with, with theft and then you distill it back.

C (21:13)

Actually this is a good point. If you go out to our website, this is one of the feature directions that we list at the very bottom. We would love to see if we could get similar performance. We do achieve state of the art performance on goal condition RL and JAX RL by a significant amount. And so it was very exciting to see the sort of frontier of the ability to train RL agents pushed. And if we can do that in a way that also sort of is just as efficient As a standard networks, that would be very cool. So is it able?

B (21:46)

Yeah, because training doesn't have to be the same thing that you deploy at inference.

C (21:50)

Right.

B (21:51)

You know what I mean?

C (21:53)

So if there's ways to distill down to a smaller model or prune the model and maybe not still retain performance, that's a very interesting research structure that we should build.

B (22:00)

Let's talk about other future directions. What else is. Is your personal passions.

A (22:05)

Yeah, so currently I'm pursuing direction of stitching in reinforcement learning. So we are trying to generalize reinforcement learning from shorter sub behaviors so that they are stitched merged during the test time. And yeah, I think this is one of my last papers that I will tackle during the PhD.

C (22:27)

Personally I would. I'm very curious of like can we like what's the like real like can we push. I'm curious about like advancing the frontier as much as possible. So if you actually look at our paper, we focus on scaling depth, but we noticed that we see that scaling width actually also improves performance. And we also find that actually by scaling depth we actually unlock the ability to scale along batch size as well. So this is one of.

B (22:51)

Yeah, so, okay, so I guess collinear, like.

C (22:55)

Yeah, right. So like, okay, I guess for context, like in traditional rl, like value based rl, scaling batch size is not super effective. But we also can see there's also other work in other areas of deep learning that show that scaling batch size is only most effective when there's like a large enough network capacity to take advantage of the scaled batch size. And we actually find that, you know, perhaps, so one hypothesis might be like, perhaps the reason why scaling batch size isn't that effective in traditional rls because like we've been using these tiny networks that haven't been able to capture that. And one of our experiments is that because we are enabled successful training of deep network, we actually were able to. This is a great test bed for testing this hypothesis. And we find that indeed as we scale to network capacity, we also unlock this different dimension of scaling by our site. And so all that to say is that I'm very curious for someone with enough compute to take some of these environments, scale up depth to the maximum capability, also scale on width, also scale on batch size. And let's like basically like in the same way that in language we're scaling on so many different axes, can we unlock different dimensions of scaling as well? And what capabilities and how far can we push the frontier of training these RL agents from doing that before we.

B (24:04)

Pass To Ishan, when you say enough compute, what kind of compute budget did you have? How does it. I just want to see what you guys got.

C (24:11)

Good question. So we wanted to make sure that this is. We wanted to make it such that you know, it's quite accessible. So the nice thing is that all of our experiments, even the thousand layer networks can be run on one single 80 gigabyte H100 GPU.

B (24:26)

So that's dollars. Yeah.

C (24:28)

Right, right, right. So everything can be run on one gpu. But in theory if we had you know like a distributed training setup and like can just like blast compute through this and really wanted to push the frontier, it'd be very interesting to see how things go. Yep.

B (24:40)

Cool.

E (24:40)

And I've actively been trying to learn as much as I can about vision language action models. Role models at Neurips. Going to a lot of machine language action models. Vision language.

B (24:49)

Vision language.

E (24:52)

And yeah curious about applications of representing robotics for these. Yeah, exactly. For robotics. Actively trying to explore more in that area. So just reading a lot of literature, talking to as many people.

B (25:02)

Yeah. We just released our episode with genuine tuition.

E (25:06)

Oh okay. Awesome.

B (25:08)

Where if you know a bit about their history, they started as a gaming clipping company and they basically have a vision language action model which I saw a preview. It was very impressive. I am not sure exactly how transferable it is to embodied use cases but it doesn't have to screen is fine. I don't know if you have any takes on.

E (25:31)

Yeah, it's an exciting research direction. Definitely.

B (25:33)

Yeah. I think the concepts of actions as something that you are outputting is actually not that popular in industry. Right. Only because text has completely dominated the last three years and tool calling which is just another form of structured text. And I feel like the action research is kind of like I don't know what needs to happen in order to unlock the next phase in that. I don't know if you've seen anything interesting I'll hear shout it out.

E (26:08)

Yeah. There's a lot of cool work on leveraging pre trained VLMs and you freeze.

B (26:14)

It and then you apply it.

E (26:14)

Yeah. And then you on top of that sort of experts to output actions. Also systems for doing hierarchical planning maybe outputting some higher level plan. And this is like a larger network that takes a long time to a little longer to do inference and so it outputs its plans with less frequency some sort of chunk and then from there there's like some sort of second system that operates a bit more fast. I think there's quite a bit of Interesting research in that direction. So that's what I'm looking forward to. Cool.

B (26:43)

Final question. Hardest question you were asked at the postal session or just favorite encounter? Anyone famous that you met?

C (26:50)

So I actually haven't gotten a chance to go to the conference that much. I'm actually working full time now, so.

B (26:54)

Oh, damn.

C (26:55)

Yeah. So so far I actually literally just got my badge like a few moments before session, so I guess I wouldn't be the best to answer that question.

B (27:04)

No, no, no. Like, you see people ask you stuff, right? Oh, that might post people asking you or meeting you and just give a vibe of what people are saying.

C (27:15)

And yeah, people were very. I think it's sort of like a very eye opening. I think that the general question is that people thought it's a very eye opening paper because like the objective is quite simple, it's quite elegant, and for us to be able to, I don't want to say overturn, but sort of challenge the conventional wisdom that RL is not super scalable and push it to such limits like 1000 layers deep and see continued improved performance. I think the general impression that I've gotten is that this could be a really cool. If we can sort of build along this direction and that we can really scale along all these different dimensions and push the frontier of the ability for rl. I'm very curious to see how that goes.

B (27:57)

All right, well, thank you so much for dropping by. Congrats on the paper again and good luck in your future work.

A (28:03)

Thank you.

E (28:03)

Thanks for having us.

B (28:04)

Yeah, thanks.