[agi] Simulation and cognition
Philip, You and I have chatted a bit about the role of simulation in cognition, in the past. I recently had a dialogue on this topic with a colleague (Debbie Duong), which I think was somewhat clarifying. Attached is a message I recently sent to her on the topic. -- ben Debbie, Let's say that a mind observes a bunch of patterns in a system S: P1, P2,...,Pn. Then, suppose the mind wants to predict the degree to which a new pattern, P(n+1), will occur in the system S. There are at least two approaches it can take: 1) reverse engineer a simulation S' of the system, with the property that if the simulation S' runs, it will display patterns P1, P2, ..., Pn. There are many possible simulations S' that will display these patterns, so you pick the simplest one you can find in a reasonable amount of effort. 2) Do probabilistic reasoning based on background knowledge, to derive the probability that P(n+1) will occur, conditional on the occurence of P1,...,Pn My contention is that process 2 (inference) is the default one, with process 1 (simulation) followed only in cases where a) fully understanding the system S is very important to the mind, so that it's worth spending the large amount of effort required to build a simulation of it [inference being much computationally cheaper] b) the system S is very similar to systems that have previously been modeled, so that building a simulation model of S can quickly be done by analogy About the simulation process. Debbie, you call this process simulation; in the Novamente design it's called predicate-driven schema learning, the simulation S' being the a SchemaNode and the conjunction P1 P2 ... Pn being a PredicateNode. We plan to do this simulation-learning using two methods * combinator-BOA, where both the predicate and schema are represented as CombinatorTrees. * analogical inference, modifying existing simulation models to deal with new contexts, as in case b) above If we have a disagreement, perhaps it is just about the relative frequency of processes 1 and 2 in the mind. You seem to think 1 is more frequent whereas I seem to think 2 is much more frequent. I think we both agree that both processes exist. I think that our reasoning about other peoples' actions is generally a mix of 1 and 2. We are much better at simulating other humans than we are at simulating nearly anything else, because we essentially re-use the wiring used to control *ourselves*, in order to simulate others. This re-use of self-wiring for simulation-of-others, as Eliezer Yudkowsky has pointed out, may be largely responsible for the feeling of empathy we get sometimes (i.e., if you're using your self-wiring to simulate someone else, and you simulate someone else's emotions, then due to the use of your self-wiring you're gonna end up feeling their (simulated) emotions to some extent... presto! empathy...). --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
Re: [agi] Simulation and cognition
Hi Ben, What you said to Debbie Duong sound intuitively right to me. I think that most human intuition would be inferential rather than a simulation. but it seems that higher primates store a huge amount of data on the members of their clan - so my guess is that we do a lot of simulating of the in-group. Maybe your comment about empathy throw intersting light on this. If we simulate our in-group but use crude inferential intuition for most of the outgroup (except favourite enemies that we fixate on!!) then maybe that explains why we have so little empathy for the outgroup (and can so easily treat them abominably). Given that simulation is much more computationally intensive it gives us a really strong reason for emphasising this capacityy in AGIs precisely because they may be able to escape our limitations in this area to great extent. AGIs with strong simulation capacity could therefore be very valuable partners (complementors) for humans. The empathy issue is interesting in the ethical context. We can feel empathy because we can simulate the emotions of others. Maybe the AllSeing AI needs to make an effort to not only simulate the 'thinking of other beings but also their emotions as well. I guess you'd have to do that anyway since emotions affect thinking so strongly in many (most?) beings. Cheers, Philip You and I have chatted a bit about the role of simulation in cognition, in the past. I recently had a dialogue on this topic with a colleague (Debbie Duong), which I think was somewhat clarifying. Attached is a message I recently sent to her on the topic. -- ben Debbie, Let's say that a mind observes a bunch of patterns in a system S: P1, P2,...,Pn. Then, suppose the mind wants to predict the degree to which a new pattern, P(n+1), will occur in the system S. There are at least two approaches it can take: 1) reverse engineer a simulation S' of the system, with the property that if the simulation S' runs, it will display patterns P1, P2, ..., Pn. There are many possible simulations S' that will display these patterns, so you pick the simplest one you can find in a reasonable amount of effort. 2) Do probabilistic reasoning based on background knowledge, to derive the probability that P(n+1) will occur, conditional on the occurence of P1,...,Pn My contention is that process 2 (inference) is the default one, with process 1 (simulation) followed only in cases where a) fully understanding the system S is very important to the mind, so that it's worth spending the large amount of effort required to build a simulation of it [inference being much computationally cheaper] b) the system S is very similar to systems that have previously been modeled, so that building a simulation model of S can quickly be done by analogy About the simulation process. Debbie, you call this process simulation; in the Novamente design it's called predicate-driven schema learning, the simulation S' being the a SchemaNode and the conjunction P1 P2 ... Pn being a PredicateNode. We plan to do this simulation-learning using two methods * combinator-BOA, where both the predicate and schema are represented as CombinatorTrees. * analogical inference, modifying existing simulation models to deal with new contexts, as in case b) above If we have a disagreement, perhaps it is just about the relative frequency of processes 1 and 2 in the mind. You seem to think 1 is more frequent whereas I seem to think 2 is much more frequent. I think we both agree that both processes exist. I think that our reasoning about other peoples' actions is generally a mix of 1 and 2. We are much better at simulating other humans than we are at simulating nearly anything else, because we essentially re-use the wiring used to control *ourselves*, in order to simulate others. This re-use of self-wiring for simulation-of-others, as Eliezer Yudkowsky has pointed out, may be largely responsible for the feeling of empathy we get sometimes (i.e., if you're using your self-wiring to simulate someone else, and you simulate someone else's emotions, then due to the use of your self-wiring you're gonna end up feeling their (simulated) emotions to some extent... presto! empathy...). --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED] --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
RE: [agi] Simulation and cognition
What you said to Debbie Duong sound intuitively right to me. I think that most human intuition would be inferential rather than a simulation. but it seems that higher primates store a huge amount of data on the members of their clan - so my guess is that we do a lot of simulating of the in-group. Maybe your comment about empathy throw intersting light on this. If we simulate our in-group but use crude inferential intuition for most of the outgroup (except favourite enemies that we fixate on!!) then maybe that explains why we have so little empathy for the outgroup (and can so easily treat them abominably). Good point. And, simulating the in-group is easier for two reasons: 1) in-group members are similar to us, so we can use our self-models as initial guesses for modeling other in-group members ... whereas if we want to model out-group members, we need to do more learning from scratch 2) in-group is often smaller than the out-group: modeling a smaller range of individuals requires less computational effort Again i come to the conclusion that the root of all evil is not money, but rather limitations on compute power... ben --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
RE: [agi] Simulation and cognition
Hi Ben, Maybe we do simulate a *bit* more with out groups than I first thought - but we do it using caricature stereotypes based on *ungrounded* data - ie. we refuse to use grounded data (from our ingroup), perhaps, since that would make these outgroup people uncomfortably too much like us. Cheers, Philip --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]