John, Actually, Nil and I outlined a mathematical theory of metalearning (in the general setting of optimization) last year, and applied it to the case of supervised classification.... The paper will get submitted for publication soon...
ben g On Wed, Sep 26, 2012 at 8:21 PM, Mike Tintner <[email protected]> wrote: > JohN:I'm surprised that you guys don't already have a mathematical structure > targeted to handle this? This is part of the core of AGI, handling the > different learning from various domains, tying that together in a model. I'm > sure you have a few candidate structures in mind... ? > Also, how could you build a general thinking machine without this > meta-learning? That, I suppose you are saying would be contained in the > particular mathematical structure used. > > Be interesting to discuss this. I hadn't come up across this category > before: > > http://en.wikipedia.org/wiki/Meta_learning > > It seems to express the profound confusion of AI about the real world > activities of real world agents - as opposed to artificial world narrow > AI's. > > The reality of human activities is that they are a continuous business of > learning, which never stops. > > And part of that learning - wh. "metalearning" may arguably overlap with - > is > > 1. learning "the rules of the game" - *as you go along*. "Learning on the > job". > > This is in complete contrast to narrow AI's which *have all the rules before > they begin*. It doesn't matter that they may still have to learn certain > aspects of a given task - they have rules for that, so they still have all > the rules before they begin. > > For example, you only learn how to play football, like all other activities, > from sex to going shopping, as you go along - you start playing before you > know more than a handful of rules, and gradually pick up (and never stop > picking up) rules, principles, strategies, new actions/skills in a lifetime > of playing. > > 2. THERE ARE NO RULES IN REAL WORLD ACTIVITIES > > Here is yet another classic case of how key concepts are used fundamentally > differently in AI and rational science from real world affairs. > > In AI, rules really are rules - they tell an agent precisely how to proceed > step-by-step. > > In real world activities, "rules" are actually only vague *principles* - and > give no indication as to step-by-step proceedings. The rules of the game in > football or tennis do not tell you exactly how to kick or hit a ball, or > exactly when or where - as narrow AI programs do. > > *No rules in the AI sense are possible for real world activities* > *No rules in the AI sense exist anywhere in our massively extensive culture > for real world world activities.* > *The only rule - it is often repeated - is that there are no rules.* > > This is also true BTW of the real world activity of programming. > > 3. Metalearning then as per John is almost certainly impossible. > > There are no general rules for conducting diverse activities - certainly > not in the programming sense. > > There are no general rules for programming. > > I'll leave it there for now. > > All comments exploring metalearning welcome. > > > ------------------------------------------- > AGI > Archives: https://www.listbox.com/member/archive/303/=now > RSS Feed: https://www.listbox.com/member/archive/rss/303/212726-11ac2389 > Modify Your Subscription: > https://www.listbox.com/member/?&; > Powered by Listbox: http://www.listbox.com -- Ben Goertzel, PhD http://goertzel.org "My humanity is a constant self-overcoming" -- Friedrich Nietzsche ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968 Powered by Listbox: http://www.listbox.com
