[hobyrne]

Jeff recognises that there is a distinction between declarative memories - stuff you can talk about, what you had for lunch yesterday, where you went on vacation - and nondeclarative memories - the example he gives is how to balance on a bicycle.

He also suggests that one of the most useful applications of intelligent machines will be just to sit and think - sensors, but no motors. This, of course, reduces fears that smart robots will take over the world. But, I hypothesise that the main (perhaps the only) reason memories get formed in a declarative manner is that we communicate. All the most important things you learn as a baby, how to move your own body, how to recognise the same object from different angles, even how to speak - the very vehicle of declarative memories - are nondeclarative. (Obviously, the declarative aspects can get hard-coded easier, the most frustrating things to AI programmers are the nondeclarative ones, by nature of their being nondeclarative.) It would seem that initially, at least, the new insights an intelligent machine may figure out about the world would be nondeclarative, just like a baby - intuitive, and inexpressible. And without observing overt communication, as a baby does, and without the means to participate, as a child has, such an intelligent machine would not have any motivation to form declarative memories, as a child does.

Finally, it seems that nondeclarative memories would be very difficult to decypher, even if we could see the entire state of the memory. If *I* can't put into words how to balance on a bicycle, even when I know how, I have little faith *you* could, even if you could look at every neuron and nerve in my body.

Conclusion: To get useful information from an intelligent machine, it will need a voice and ears (or equivalents, the point being bidirectional communication with people, possibly requiring the output to be fed back to the input - babies learn to speak by detecting how the sounds it makes are different from the sounds others make), and training in standard ways to express ideas that we humans can understand. Such machines will go through a phase of infantile jabber, and baby-talk, before they become usefully communicative. Even so, it may not be able to describe its most fundamental insights, as we cannot describe many of ours in a declarative way.

[skilesare]

I'm right there with you. I'm working on a model that adds this layer. I've got some ideas about how it will work. I'll post my results as I get them.

There is a whole lot that I don't understand yet, but I think I'm getting close mimiking what Dileep has done in his image experiments.

[tos]

Despite many attempts (my previous entries seems to have been erased or lost): I do not agree.
It is perfectly possible to make Knowledge based systems that can learn how to balance on a bicycle, for example.
We do that in a newly formed company in Norway.
_________________
-----
Kind regards, Tos
‘To punish me for my contempt for authority, fate made me an authority myself.’ - Albert Einstein

[DavidOlmsted]

I think a better word for declaritive memories is "episodic" which seems to be used more often by psychologists and neuroscientists. Episodic memory is that memory which is formed from experienced events in space and time. They occur in episodes clustered around certain key (big or emotional) events.

In that sense Hobryne is correct. Many other forms of learning exist besides episodic memory, the most important is the adaptive feedforward learning for skills which is independent of the time or place where it was learned.

I suspect the memory used in knowledge based systems for control is more analagous to skill based memory.

Yet episodic memories can transform into habits like skill based memory so I would not want to get bogged down in definitions until we actually simulate the brain to see what is there.
_________________
Click to go to my site at neurocomputing.org

[tos]

Take a look at www.alfatroll.com
_________________
-----
Kind regards, Tos
‘To punish me for my contempt for authority, fate made me an authority myself.’ - Albert Einstein

[DavidOlmsted]

TOS, from your website it seems you are doing some interesting work although judging its capabilities is not possible until you release more information on the processes involved. I understand you may not want to do that due to patent and competitive concerns but I would be interested in reading what you can make public.

Resonance matching sounds like you are using some sort of analog value to find the "best fit" instead of conventional logic.
_________________
Click to go to my site at neurocomputing.org

[tos]

Hi again.
It has become more clear now that the technology presented at www.alfatroll.com is a reasoning engine, suited for autonomous control in servers, embedded systems, as well as for other programming purposes as well.

I have also become aware of the fact that the simplicity of the method, and its predictability re. response times may make it suitable for certification as an onboard system in civilian airspace.

No "reasoning engine" up to now has been certified for that purpose, and I would be grateful for tips or contact persons for entering a registration process.
_________________
-----
Kind regards, Tos
‘To punish me for my contempt for authority, fate made me an authority myself.’ - Albert Einstein