I do not remember hearing of transfer learning before although I have thought about things like that. I am still a little skeptical about deep learning because there has been such a lack of obvious cross-application. As I understand it, transfer learning can be applied when there are a number of significant features in a trained model which are also signnificant to another model or problem. This is something I would like to try but I don't really have time to do so. The problem with deep learning, as I see it, is that a DL model can be trained but it does not have insight into the very features that it is trained to detect. (That does occur with naïve students.) And even if transfer-learning was used the result would not be a single integrated DL model but numerous discrete DL models. This is OK but I believe that it is evidence that AGI - like knowledge has to be a hybrid of network learning and discrete learning. So I am really thinking of a discrete network kind of AI where 'transfer learning' could take place more automatically. In the kind of discrete network that I have in mind 'transfer learning' would be an integrated part of all learning. The program would be looking for features in the data which could be found and used in numerous ways. The problem with this idea is that there are so many different ways to group individual data points (into potential groups and generalizations) that the system would be overwhelmed at the start with a combinatorial explosion of possible groupings or generalizations. This combinatorial explosion is the complexity problem. So a neural network has one way to simplify the problem but it is not the only possible way to do so. I think this is a serious problem and the difficulty of using a trained DL model to recognize the individual features that occur in the data that contains something that it is trained to detect is why DL is not AGI. It is not even narrow AGI. It may become narrow AGI but it definitely is not there yet. Jim Bromer
On Thu, Aug 8, 2019 at 12:46 PM Brett N Martensen <[email protected]> wrote: > Jim, You are right on the money! > It's called transfer learning and comes from having generalization in a > compositional hierarchy in which more complex things are composition of > simpler but more general things. And the lowest level simplest things yet > most general are the stimuli that come from sensors and that also makes it > grounded. > Brett > *Artificial General Intelligence List <https://agi.topicbox.com/latest>* > / AGI / see discussions <https://agi.topicbox.com/groups/agi> + > participants <https://agi.topicbox.com/groups/agi/members> + delivery > options <https://agi.topicbox.com/groups/agi/subscription> Permalink > <https://agi.topicbox.com/groups/agi/T187cd9f14076b86f-Mdd205164b317e14857615940> > ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T187cd9f14076b86f-Md7a72ea7b52e0eea42e4706f Delivery options: https://agi.topicbox.com/groups/agi/subscription
