> By the way:
> I'm going to use BPTT (backpropagation through time) to train the recurrent
> neural network. According to Schmidhuber at IDSIA there is a better
> recurrent neural network system Long Short Term Memory (LSTM). This network
> + training technique is claimed to be able to look further back (>1000
> steps in stead of just 10 for BPTT). Is anyone familiair with this? Can it
> be put in the same layering architecture that I am aiming at, i.e. does it
> have an abstraction state?
>From what I've read about prediction with recurrent neural networks (and
similar systems), it looks to me like only systems that predict detail
information are investigated. The main AI strategy there is to make systems
that predict detail states as far as possible: 100 steps in the future, 1000
steps... 10000... But I don't think that will ever go so far as to make long
term prediction and planning possible. In a realistic environment a system
will never be able to predict all the detail states the system will be in
over the range of a year or so.
A system will at most be able to predict in which class the detail information
will fall. And the farther into the future the prediction, the larger such a
class will tend to be, i.e. the prediction will be in more general terms.
There are symbolic reasoning systems that try to plan globally in general
terms. But with symbolic computation the problem is that the symbols lack
semantic content. There is no or are only bad connections to perception and
action and different layers of abstraction. These systems just reason in
their own private syntactic Language of Thought, in an autistic way. Or they
can only act successful in simple 'toy' worlds.
A reason I can think of for not trying to predict the internal state of a
recurrent neural network is that then a prediction is made about an internal
construction, not something external, and furthermore this prediction has no
further effects (for prediction). It stays internal. The prediction of the
internal state of the recurrent network is seen as a goal on its own. It is
not related to predicting the detail level better.
This is counterintuitive. If it doesn't help at the detail level then it seems
useless. However, the internal state is determined by the external detail
level (the perceived sequence). And therefor, indirectly prediction of the
internal state contains information about the future of the external
environment.
The building up of knowledge, concept formation and abstraction, when there
are multiple recursive layers of internal state predicting (of a lower layer)
recurrent neural networks, can be the input for action systems (also neural
networks) at each layer. The actions at the higher levels are abstract plans
or goals (formulated in the representation of the layer below) spanning a
long range of time (the typical time range of the predictions of that layer).
The action system passes such a plan or goal to the layer below where it is
the input for the action system there. The goals/plans of the lower levels
become more concrete and short term. At the bottom the action is really
concrete, acting in the environment through effectors ('muscles'). (All
action systems try to maximise an external reinforcement signal through their
actions.)
So this is how the high level abstract representation states have their effect
on the detail level; through action systems, the results of which are passed
down via via to the detail level.
Arnoud
-------
To unsubscribe, change your address, or temporarily deactivate your subscription,
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
