* Boston Dynamics engineers create physical models Why do typos mostly only show up after pressing the submit button? :)
On Monday, February 8, 2021 at 3:46:31 PM UTC Patrick Hammer wrote: > Hi Jose! > > Thank you for initiating this interesting discussion! > I guess there are truths in both Sutton's and Brooks views, as often in AI > the reality lies somewhere between the extremes! :) > Undoubtedly Deep Learning has made obsolete for instance the comparably > way less fruitful approach of feature engineering, here I agree with Sutton. > On the other hand, Brooks has correctly identified that human expertise is > now utilized in the design process of the layers, models and loss functions > before their parameters are actually optimized. > > Personally I'm quite agnostic to whether "human engineering" versus > "offline-optimization within human-defined boundaries", is better, both are > just two different paradigms of engineering which can also be combined. > While offline-optimization (via Supervised DL especially) has taken over > in many domains, for some cases explicit engineering is still superior. An > example are the famous legged Boston Dynamics robots: Boston Dynamics > engineers physical models and throws them into Model Predictive > Controllers, instead of applying any Reinforcement Learning. While there is > plenty of research in using Reinforcement Learning in legged robots (often > in a RL&Control hybrid approach), these solutions don't perform comparably > well so far. Part of the reason is that offline-optimization demands an > accurate simulation to work out. This is clearly the case for computer > games and board games (perfect simulation availability even there), but not > so well for systems which need to operate in the real world! > > What matters to me personally is not the particular engineering paradigm > to create systems for a specific purpose (via offline-optimization and > handcrafting), but whether the AI can effectively adapt, at runtime, to new > circumstances. That's a big challenge, and is what distinguishes, at a high > level, natural evolution from natural intelligence (whether a single > individual can adapt, or whether multiple generations are necessary). Most > AGI systems, including OpenCog Prime address this quite well in my opinion, > and realizing that's the case was a large part of why I was pulled into > this wonderful research field! > Recently, our team has also written a Blog post on this topic, which also > addresses the "Generality vs Specialization" issue you have touched on: > http://www.opennars.org/blog/post1.html > > Best regards, > Patrick > > On Wednesday, November 11, 2020 at 1:42:43 AM UTC Jose Ignacio > Rodriguez-Labra wrote: > >> There was an earlier thread about Sutton's bitter lesson (link >> <http://incompleteideas.net/IncIdeas/BitterLesson.html>), which >> basically argues that general machine learning methods are always better >> than specialized methods encoded with human knowledge and optimized, which >> seemed like most people agreed with. There is a response on it called The >> Better Lesson by Rodney Brooks (link >> <https://rodneybrooks.com/a-better-lesson/>), pointing out reasons why >> Sutton is wrong. I really recommend giving it a read. >> >> It made me think about how using certain concepts that we already know >> about the world could actually be useful, rather than building a completely >> blank environment and have it learn everything from scratch. Why throw away >> all the patterns we've recognized already? Plus we can't rely on the >> increase in compute (link >> <https://venturebeat.com/2020/07/15/mit-researchers-warn-that-deep-learning-is-approaching-computational-limits/>), >> >> which is integral to general methods, and playing into the process >> perpetuates the ever-increasing carbon footprint of the machine learning >> industry. >> >> There seems to be a duality between these two methodologies: generality >> and specialization. Which is the right approach? But they could work >> together. By using our human ingenuity and our current understanding of the >> brain, maybe we could build a specialized, but limited version of human >> intelligence, to then use to create a general intelligence. Perhaps a truly >> general method for building human intelligence is a task belonging to a >> post-singularity world. How else could we overcome such a large problem >> space? >> >> What do you think? is there any merit to this,? Or I am just not >> experienced enough? >> Maybe I should stop thinking so much and get coding. >> > -- You received this message because you are subscribed to the Google Groups "opencog" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To view this discussion on the web visit https://groups.google.com/d/msgid/opencog/c843cab6-d4d9-4642-9c97-59c2b160beddn%40googlegroups.com.
