Matt Mahoney <[email protected]> wrote: It's called "wishful thinking". He proposes "cognitive synergy" as an excuse for not testing. When all of the components are put together, it will magically work. It's just intuition, of course, not backed by any evidence. In fact, all of the evidence points the other way. The most powerful models in machine learning are ensemble models. You combine lots of predictors and get more accurate predictions. If you remove half of them, then you still get most of the accuracy. Each model can be tested independently of the others, because that's how they work in practice. Some examples: - Watson is made up of hundreds of independent modules. Each is able to answer a small subset of Jeopardy questions.
That is interesting. There should be some modularization in early AGI, but not the way you describe it here. The reason is because people will not be able to engineer the device until someone figures out a working solution to the fundamental problems. That means that it is very likely that key parts of the program will be haphazardly distributed throughout the program. Typically, in a situation like this, you will find that the program seems to be working for a little while and then surprisingly does not generalize (or continue) for reasons that you cannot even guess at. But then all of a sudden, it will start working again for a short period before sputtering out. A better model for testing is to demonstrate that some key abstract problems can be overcome and then, by working on these, come up with a way to show that it can be generalized to other problems of the same kind and to problems that are related but essentially different in someway. Then you have to keep going, you have to keep showing progress. Jim Bromer On Wed, Apr 10, 2013 at 10:00 PM, Matt Mahoney <[email protected]>wrote: > On Wed, Apr 10, 2013 at 6:29 PM, Tim Tyler <[email protected]> wrote: > > > > Practically any form of software development involves lots of testing. > > > > However, looking at: > > > > > http://multiverseaccordingtoben.blogspot.com/2011/06/why-is-evaluating-partial-progress.html > > > > ...certainly suggests that Ben has some rather odd ideas about testing. > > It's called "wishful thinking". He proposes "cognitive synergy" as an > excuse for not testing. When all of the components are put together, > it will magically work. It's just intuition, of course, not backed by > any evidence. In fact, all of the evidence points the other way. The > most powerful models in machine learning are ensemble models. You > combine lots of predictors and get more accurate predictions. If you > remove half of them, then you still get most of the accuracy. Each > model can be tested independently of the others, because that's how > they work in practice. Some examples: > > - Watson is made up of hundreds of independent modules. Each is able > to answer a small subset of Jeopardy questions. > - The PAQ compressor is made up of hundreds of independent bit predictors. > - People partially recover from strokes because other parts of the > brain compensate for the parts that are destroyed. > - Stephen Hawking and Helen Keller are missing some key components but > are still considered intelligent. > > Early progress in AI was rapid, but then stalled after we solved all > the easy parts. You can't blame this on cognitive synergy. If it were > true, then progress would have been slow at first, then picked up > speed as we got closer to the finish, the opposite of what we > observed. > > It looks to me like OpenCog is going the way of NARS. You may recall > how how Pei Wang spent over a decade developing a general data > structure for knowledge representation and a mathematical model of > learning and reasoning. It has many of the same elements as AtomSpace: > truth values, confidences, is-a links, logical operations, etc. But it > ended up going nowhere. He never did any of the hard work of > collecting training data and testing it on real-world problems like > text prediction or image labeling or robot navigation. He never > estimated how much data he would need, or how much computing power to > process it. > > As you read this, your brain is computing 10^15 weighted sums per > second on 10^14 weights, and then adjusting them according to a > complex algorithm that depends on 3 x 10^9 DNA bases, equivalent to > 300 million lines of code written by a 3 billion year long search > algorithm running on a planet sized molecular computer. Maybe there is > a way to do this on your PC, but I really doubt it. > > I am not saying this because I want to see OpenCog fail. I would > rather see research and get answers to hard questions. There is a lot > we still don't know. > > -- > -- Matt Mahoney, [email protected] > > > ------------------------------------------- > AGI > Archives: https://www.listbox.com/member/archive/303/=now > RSS Feed: https://www.listbox.com/member/archive/rss/303/10561250-470149cf > Modify Your Subscription: > https://www.listbox.com/member/?&; > Powered by Listbox: http://www.listbox.com > ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
