BAIR Blog

The article doesn’t provide any alternatives? I think there are indeed many challenges when evaluating Compound AI Systems (http://bair.berkeley.edu/blog/2024/02/18/compound-ai-systems/) But evals in complex systems are the best we have at the moment. It’s a “best-practice” just like all the forms of testing in the “test pyramid” (https://martinfowler.com/articles/practical-test-pyramid.html) Nothing is a silver bullet. Just hard won, ideally automated, integrated quality and verification checks, built deep into the system and SDLC.

Forgot to share this link as well, not sure if you're aware of it but it's a great write up on fine tuning small local models on specific APIs and seems like it would be a perfect fit for your project. https://bair.berkeley.edu/blog/2024/05/29/tiny-agent/

I don't think people have given up on model based RL, it is just that describing a proper model is (like you are saying) very difficult. in the case you haven't seen or read the following: https://bair.berkeley.edu/blog/2019/12/12/mbpo/

This is great stuff. I actually think the obvious application here is foreign language instruction. Particulary chinese character recognition for english speaking elementary students. Most teachers at that level will probably not have mastery of the language themselves. In addition, the "teacher agent" may be able to find patterns that would dramatically boost recall. Additional meta-learning research from openai posted recently on arxiv: Evolved Policy Gradients https://arxiv.org/pdf/1802.04821.pdf As well as the Ray distributed AI system: http://bair.berkeley.edu/blog/2018/01/09/ray/

The inaugural post is worth reading: http://bair.berkeley.edu/blog/2017/06/20/learning-to-reason-with-neura… It explores a combination of neural networks and structured parsing to solve challenges at the intersection of vision and natural language processing.

Just some context here: * The raw data in nuscenes ( https://www.nuscenes.org/ ) is about 5x larger than this dataset from Lyft. 300GB train vs 60GB train. Argoverse ( https://www.argoverse.org/ ) is is about 3.x larger at 200GB. The Waymo dataset will (allegedly) be an order of magnitude larger than nuscenes ( https://i2.wp.com/syncedreview.com/wp-content/uploads/2019/06/image-55… ). BDD100k ( https://bair.berkeley.edu/blog/2018/05/30/bdd/ ) is the "largest" public dataset to date, but lacks lidar, and labels are inconsistent; most of the 100,000 scenes only have one labeled frame. * The Lyft sensor suite has bumper-mounted lidar, which is absent from other existing datasets. Point cloud data in these areas is critical for pedestrians, bikes, and various road hazards. So this dataset alone is useful for validating work trained through other means. * The current Lyft Level 5 release has no explicit test / validation set, which is crucial for properly measuring performance of any experiment one might do with the data. In nuscenes and Argoverse, there's a small snippet dataset that helps you prepare your pipeline. Feels like Lyft might have rushed things a little here-- they could have posted a "teaser" and then the full train and test/validation set a couple weeks later. Great to see more public data (especially from a more modern sensor suite), plus investment into a contest with prizes.

This is why the "AI learns from materials just like a human does so it's not copyright infringement" argument always bothered me. A person won't recite full pages of word-for-word copies [1] from their head when you ask them something. When I first tried Copilot, I asked it to write a simple algorithm like FizzBuzz and it ripped off a random repo verbatim, complete with Korean comments and typos. Image models will also happily generate near-identical [2] copies (usually with some added noise) of copyrighted images with the right prompt. [1] https://bair.berkeley.edu/blog/2020/12/20/lmmem/ [2] https://www.theregister.com/2023/02/06/uh_oh_attackers_can_extract/

Very Happy to see FIGS on the list. I almost got a chance to use it on a recent project but the csutomer in the end decided they valued accuracy over interpretability. I wonder if the recent KANs fit into this? They're more interpretable than other NN architectures. Also, Berkeley has had a few articles on interpretable methods (including FIGS) over the years but idk how they compare to what you've already implemented: https://bair.berkeley.edu/blog/2022/02/02/imodels/ https://bair.berkeley.edu/blog/2020/04/23/decisions/

You could possibly fine tune it for narrow domain tasks like they did with tiny-agent https://bair.berkeley.edu/blog/2024/05/29/tiny-agent/ I like the approach that Apple seems to be taking with fine tuned small models that handle routine tasks and then defer to larger off device models for things they can’t confidently do. I imagine you could construct a training set that contains examples that should produce low confidence answers where you could add an output that is essentially a “call for help” option so you could train it to choose that. Smaller models also means you could have more running in parallel and use another to route requests to the appropriate expert.

Glad to see MLC is becoming more mature :) I can imagine the unified engine could help build agents on multiple devices. Any ideas on how those edge and cloud models collaborate on compound tasks (e.g. the compound ai systems: https://bair.berkeley.edu/blog/2024/02/18/compound-ai-systems/)

I wouldn't say it is an unexpected behavior. I remember reading papers about this memorization behavior few years ago (e.g., [1] is from 2019 and I believe it is not the first paper about this). It should be expected from OpenAI to know that LMs can exhibit memorizing behavior even after seeing the sample only once. [1] https://bair.berkeley.edu/blog/2019/08/13/memorization/

Not sure why this post is getting down voted? It's spot on - the limitation lies in relatively small text encoders within stable diffusion/mid journey. Larger language models contribute to better image generation (see: DeepFloyd). Google's CapPa also recently showed an alternative to the contrastive learning of CLIP at any param count. Until larger models are released, there are "hacks" like this available: https://bair.berkeley.edu/blog/2023/05/23/lmd/ - GPT4-generated bounding boxes guiding stable diffusion

How is it loose? Both in the colloquial sense and in the sense it is used in machine learning it is fitting. https://bair.berkeley.edu/blog/2020/12/20/lmmem/ is a post demonstrating it.

Project Ray is specifically more focused on scaling AI systems it seems. That is sort of important but also missing from the article. https://bair.berkeley.edu/blog/2018/01/09/ray/

In our research, we show that the fovea observed many times in biology may be an emergent property that emerged from an attention system: https://bair.berkeley.edu/blog/2017/11/09/learn-to-attend-fovea/

See also Trevor Darrell's group's work on neural module networks: https://bair.berkeley.edu/blog/2017/06/20/learning-to-reason-with-neur…

For more information about why meta-learning (and in particular meta-RL) is useful see this post: https://bair.berkeley.edu/blog/2017/07/18/learning-to-learn/

We just revealed this to the world today! Excited to bring our research to a general audience in a more fun to read format (and if you have not seen it, be sure to check out the BAIR blog https://bair.berkeley.edu/blog/).

Sounds close to what Berkeley AI and Robotics lab is developing with model agnostic meta-learning. pytorch implementation here: https://github.com/tristandeleu/pytorch-maml-rl And demo where robotic arm learns to navigate a simple task after watching a single video of human doing same ;) https://bair.berkeley.edu/blog/2018/06/28/daml/

Isn't this like an "auto-regressive feed forward neural net"[0]? > Instead of making predictions from a state that depends on the entire history, an autoregressive model directly predicts yt using only the k most recent inputs, xt−k+1,…,xt. This corresponds to a strong conditional independence assumption. In particular, a feed-forward model assumes the target only depends on the k most recent inputs. Google’s WaveNet[1] nicely illustrates this general principle. [0] http://bair.berkeley.edu/blog/2018/08/06/recurrent/ [1] https://arxiv.org/abs/1609.03499

this looks promising https://www.youtube.com/watch?v=Ul0Gilv5wvY there is also this http://bair.berkeley.edu/blog/2018/04/10/virtual-stuntman/

I have to assume you're not familiar with his lab at all then: http://bair.berkeley.edu/blog/?refresh=1 I mean one of their papers on the stuntmen was on the front page of HN recently. They are most certainly on the forefront.

Related article on similar hierarchical / compositional policies learned by maximum entropy optimization: http://bair.berkeley.edu/blog/2017/10/06/soft-q-learning/

You're in the middle of a saddle shape (see the figure [1]). This is bad because you could be lower (the goal is to be at the lowest point). Moreover, getting out is apparently slow / difficult. The authors explain "how to escape" the saddle point in an efficient way. [1]: http://bair.berkeley.edu/blog/assets/saddle_eff/strictsaddle.png

Some are, some aren't. See Koala for instance. The problem with Koala is that it fine-tunes on open sourced data, but makes no claims about the data for the base LLaMA models. https://bair.berkeley.edu/blog/2023/04/03/koala/ The irony is that openAI and Meta themselves might be in flaky ground for having trained models on other people data with dubious rights to do so in many instances, and then using it to produce output commercially. But this is a new frontier and enforcement might be effectively not possible unless new legislation requires reproducibility and audits on the data sets or something like that. But without that, how do you know exactly how did they arrive at a given set of weights with Montecarlo algorithms and arbitrary fine tuning? You basically don't know what was there and you cannot prove they didn't achieve those results with perfectly clean data. PS: https://medium.com/geekculture/list-of-open-sourced-fine-tuned-large-l…

There are tons of research projects that do exactly what you're talking about, initializing an agent with zero knowledge in an environment which just provide rules of discovery and reward. The agent then takes actions in that space and learns from its own experiences. It's just that it is hard to compare with humans, who have had the benefit of evolution over millions of years. A basic example is that of AlphaGo Zero[1], which learns to play a top level game by being given enough time and just the rules of the game. This is similar to how a child learns to walk, it is just that we only have the capability to model toy situations (a board game in this instance) right now, and access to harder instances (movement in the real world) will slowly come about. There are cases of robots being programmed to poke / move / pick objects to try and learn about their shapes[2], in case you are interested in another such example. [1] - https://en.wikipedia.org/wiki/AlphaGo_Zero [2] - https://bair.berkeley.edu/blog/2019/03/21/tactile/

I found a few more, if you want a page which aggregates all of the RSS feeds, look here: https://news.aiapipro.com/ Feeds that I found: - https://openai.com/blog/rss.xml - https://hnrss.org/newest - https://towardsdatascience.com/feed - http://feeds.feedburner.com/blogspot/gJZg - https://machinelearningmastery.com/feed/ - http://distill.pub/rss.xml - https://bair.berkeley.edu/blog/feed - http://karpathy.github.io/feed.xml - http://fastml.com/atom.xml - https://aws.amazon.com/blogs/machine-learning/feed/ - https://www.analyticsvidhya.com/blog/feed - http://news.mit.edu/rss/topic/artificial-intelligence2 - https://www.marktechpost.com/feed/ - https://deepmind.com/blog/feed/basic/ - https://qudata.com/en/news/rss.xml - https://machinelearning.apple.com/rss.xml - https://www.oreilly.com/radar/topics/ai-ml/feed/index.xml - https://blogs.rstudio.com/ai/index.xml

Interesting. Some highlighted links from the writeup: 1. Differentiable plasticity: training plastic neural networks with backpropagation (https://arxiv.org/abs/1804.02464) 2. Born to Learn: the Inspiration, Progress, and Future of Evolved Plastic Artificial Neural Networks (https://arxiv.org/abs/1703.10371) 3. Github for the project: https://github.com/uber-common/differentiable-plasticity 4. Learning to Learn (http://bair.berkeley.edu/blog/2017/07/18/learning-to-learn/) 5. Meta-Learning: http://metalearning.ml/