3 ms·
"Emergence of the generally agreed upon "next big thing" in AI beyond deep learning. ... What we are seeing is new understanding of capabilities missing from th
by Animats 4y ago
"Emergence of the generally agreed upon "next big thing" in AI beyond deep learning. ... What we are seeing is new understanding of capabilities missing from the current most popular parts of AI. They include "common sense" and "attention". "
Right. As I've said on here a few times over the last few years, the big unsolved problem ahead is common sense, defined as getting through the next 30 seconds of life without screwing up. There's been too little progress on this. We need this more than ever, because we now have machine learning systems based on optimization which do the right thing most of the time and something badly wrong some of the time. (I used to speculate on something that made a sheaf of predictions and looked for good and bad outcomes, as an approach to dealing with locomotion over rough terrain. You want to pick moves which have some good outcomes and no really bad fall-off-cliff outcomes. But I was in that too early, in the 1990s, when we didn't have enough compute power for wasteful approaches like that.)
"Despite some impressive lab demonstrations we have not actually seen any improvement in widely deployed robotic hands or end effectors in the last 40 years"
Yes. Manipulation in unstructured environments is still very poor. Brooks' startup in that area, Rethink Robotics, failed on this problem. Mediocre bin picking robots were available in the 1980s, and slightly less mediocre bin picking robots are available today. Amazon has put considerable effort into this without getting good enough performance to deploy.
I had hopes for Willow Garage, which got far enough to fold clothing, but they didn't get beyond that point. It's embarrassing to compare the 1960s video from the Stanford AI lab of a robot assembling some automotive parts with the videos of the DARPA manipulation challenge from a few years ago.
"An AI system with an ongoing existence (no day is the repeat of another day as it currently is for all AI systems) at the level of a mouse."
Yes. I've been arguing for mouse-level AI, or sometimes squirrel-level AI, for a long time.
Small mammals have very good locomotion and manipulation, with brains 1000x smaller than human.[1] We're there on having enough compute power for that size of brain.
That's an indication that this problem requires a different approach, not just bigger models.
I had this argument with Brooks decades ago, when he was promoting Cog and trying for human-level AI in one jump. He said "I don't want to go down in history as the man who developed the world's greatest artificial mouse". Then he went on to develop the Roomba.
[1] https://en.wikipedia.org/wiki/List_of_animals_by_number_of_neurons https://en.wikipedia.org/wiki/List_of_animals_by_number_of_n...
- hbarka 4y agoI once heard a professor say that common sense is knowing when to make an exception. I wonder if this goes against the grain of machine learning systems which seeks to optimize the opposite of the exception.
- Animats 4y agoIntrospective definitions are not that useful. That's why I prefer "not screwing up badly" as the key property of common sense. It can be applied to observed behavior, animal or human. It's about having some way to predict the likely results of actions before doing them, and using that information to help choose actions.
- hbarka 4y agoAn anecdote was shared about a hospital janitor who was admonished for not mopping a patient’s room during the cleaning scheduled hours and it turns out this janitor was considerate about the patient’s family visiting their child, and therefore was doing the cleaning during the off-hours. The janitor was exercising common sense in consideration and empathy for the family. If the heuristic is not to screw up badly then we would all be selfish.
- ripe 4y ago> Regarding this point: "Despite some impressive lab demonstrations we have not actually seen any improvement in widely deployed robotic hands or end effectors in the last 40 years" : Yes. Manipulation in unstructured environments is still very poor. Brooks' startup in that area, Rethink Robotics, failed on this problem. (I worked at Rethink Robotics) Rethink Robotics was not working on the problem that Brooks is complaining about here. Today’s robots are still using parallel electric fingers or vacuum suction cups, which are much poorer than even the most primitive claws in the animal kingdom. That’s the problem.
- Animats 4y agoMany robot hands have been built. These are rather nice.[1] But they're usually used for teleoperators, not computer-controlled. Usefully controlling them in unstructured situations is still hard. [1] https://www.shadowrobot.com/ https://www.shadowrobot.com/