Re: Can operant conditioning account for all learning?

On Mar 29, 4:06 am, c...@xxxxxxxx (Curt Welch) wrote:
The brain doesn't store raw sensory data. It doesn't
even attempt to do it. There's way too much of it.


And I believe that is understood. I don't believe
everything meant all the raw data.


It's a large associative memory system which stores
behaviors - it stores in its wiring how it will react
to a given class of sensory stimulations.


This is something you need to flesh out.

What is a "behavior" and a "class of sensory stimulation".
If you want to actually build machines based on such
generalities you need to start fleshing out what you
actually might mean by this in actual hardware/software.


When we have what we call a "memory" of an event, we
are not replaying the raw sensory data form that event.
Instead, the brain is reproducing part of the reaction
we had to that event. We remember not the raw sensory
event itself, but we remember some aspect of how we
react to it. That's what memories are - it's the brain
reproducing a past behavior.


Ok. Flesh out what you mean by "some aspect of how we
react to it". What aspect exactly? How is this aspect
extracted?

And then you change that to "reproducing a past behavior".
What constitutes a "behavior" exactly in this context?

Now I can feel a "you lack powers of abstraction" accusation
coming on but the reality is I can wave my hand about making
high level abstractions but until you tease it all out and
clarify the vagueness and generalities of the statements such
abstractions are just wishful notions.

You say you want to work on ideas for implementing your hand
waving ideas but that requires getting some kind of handle
on those ideas, a handle that lends itself to suggestions as
to what to implement.


For example:
... it needs the power to react to temporal patterns - how
things change over time.


Yes planes need the power to fly. Got that. But it doesn't
help in inventing planes that fly to say they need the
power to fly.

Let's tease out what constitutes a temporal pattern, how it
is represented in the system, and what it means to "react".
Sure I know what all that means as a high level abstraction
but when you start trying to tease it out suddenly you find
problems your high level abstractions didn't recognize.


In the case of the brain, there seems to be a system in use
which allows it to keep higher resolution data about more
recent events with less and less memory allocated to the
data the more time passes.

I would suggest it uses the same amount of memory more
efficiently. Most likely the brain doesn't store things in
a list the way we might in a GOFAI program which has the
problem of the list getting larger and larger and any serial
search of the list taking longer and longer. I think you
have to implement an associative like memory (where the
address is the query) and a hierarchical like structure.

I see us building up a model of the world without the need for
the reward system, which we can refer to, to understand our
current input. We can also build up memory of the events in
that model like a story of our world and our part in it.
We have general facts (trees), specific facts (this tree) and
episodic facts (I was born in America, my parents were Jack
and Jill, I went to school in New Jersey .... and so on).

With damage to the hippocampus you can lose the ability to
keep adding to this story but you do not lose the model of
the world and the story before the damage. Nor do you lose
the ability learn new physical skills, like riding a bike,
which flags a different memory system is involved. Damage
to another part of the brain may keep your story but damage
the world model.


But in no case, is there an attempt to remember all the
raw sensory data.


I don't believe that was ever implied.


If you had more memory available for short term memory, it
would be used to store higher resolution information, for
longer periods. It in effect uses techniques to find, and
store only the most valuable data in the stream and to throw
out the rest. With more memory available, you could use
more, and/or higher resolution sensors to allow to you find
and store more useful data. But never would memory become
infinite and never would you be able to store everything you
would like to store and collect if you could.


I think all that is very well understood.


JC
.

Relevant Pages

  • Re: Exchange 2003 Private Store Corruption
    ... Once you've pulled the faulty RAM, create the new mailbox store and move ... then you can start working on the problem mailboxes. ... have DIMM's installed in pairs in this server? ... Online Diagnostic Utility says "Correctable Memory Threshold Exceeded" ...
    (microsoft.public.exchange.admin)
  • mm/ hacks -- Generic memory store?
    ... I want to do a memory management hack for generic memory stores. ... "store" for excess memory before OOM. ... RAM swap ... store would be capable of providing compression or encryption. ...
    (Linux-Kernel)
  • [PATCH V0 0/4] kztmem: page cache/swap compression via in-kernel transcendent memory and page
    ... The data store for this page-oriented data, ... independent data sources to share PAM resources effectively ... Transcendent memory, or "tmem", is a clean API/ABI that provides ... The first tmem implementation is in Xen. ...
    (Linux-Kernel)
  • Re: Exchange 2003 Private Store Corruption
    ... Information Store First Storage Group: The database page read from ... Online Diagnostic Utility says "Correctable Memory Threshold Exceeded" ... the server back up and running without the faulty memory installed. ...
    (microsoft.public.exchange.admin)
  • Re: Same question - Why use a DataTable in ASP .NET?
    ... When a good deal of memory is consumed, ... For small web applications that don't get ... once you start thinking "high load" ... other words, if you have a primary key, store it in the session, not ...
    (microsoft.public.dotnet.languages.csharp)