Re: Still waiting for a solution.

J.A. Legris wrote:
On Apr 13, 3:10 am, Sanny <softta...@xxxxxxxxxxx> wrote:
True, just doing arithmetic is easier, but I think learning how to do it
is a demonstrationAIproblem. Humans learn addition. It seems to me
that figuring out how to build anAIsystem whose behaviours can be
shaped so that it can add/subtract/etc would be very instructive.
So I think OP is onto something: ANN systems that can learn to do what
other systems can be programmed to do are a test of both ANNs andAI.
I want to know what type of Nural Network can perform Addition/
Subtraction/ Multiplication with 90% acuracy after feeding them with
1000s of data.

This will be easy to test whether the Nural network is working
correctly or not.

Still my Question is Unsolved.

Bye
Sanny

Start with 1-digit decimal addition: sum = X + Y + carry

Input layer: 10 units for X (0 to 9), 10 units for Y (0 to 9), 1 unit
for the carry input.
Hidden layer: 20 units for the sum (0-19)
Output layer: 1 unit for the carry output and 10 units for the 1-digit
sum.

Train 1 network using backpropagation. When it works satisfactorily,
construct and chain together N-1 identical copies for N-digit
addition.

--
Joe


So what you get an ALU -- in fact, as I understand it, you've just translated an ALU into a an ANN. That's not what I think of as a device that "learns to do arithmetic."

OTOH, what you've described is not learning, but the acquisition or development of a brand-new behaviour (1). It seems to me that your device may be a good model of development, in that the ANN's electronics may be seen as its genetics, the input as the environmental stimulus, and backpropagation as the genetic program that restructures the ANN. That is, if you accept development as the action of genes in response to environmental control (external input).

(1) I find it odd that so many AI workers seem to believe that reinforcement learning (conditioning) is exemplified in blank-slate machines. There is no blank slate. Thus a machine that "learns arithmetic" must be one that has some pre-existing behaviours which are shaped to produce the behaviours we call "doing arithmetic." The trick is to figure out what pre-existing behaviours could be thus shaped. I think Glen's exposition is a start.

Fact is that many animals can distinguish between large and small groups of objects. This behaviour can be refined in some of them (eg, parrots). NB that in these cases, we have a pre-existing behaviour that is shaped by the trainer (which could of course be the habitat of the animal.) It may be that this is a prerequisite for learning to add. I don't know.

--
wolf k.
.

Relevant Pages

  • Re: Are neural networks worth the pain.
    ... How long should I continue will the Nural network learn the Addition. ... So where do those behaviours that are "successive approximations" come ... an ANN can't do it. ...
    (comp.ai.philosophy)
  • Re: Implausibility that we can be explained by evolution
    ... the ANN in the robot can be explained to work, ... because the mechanisms and behaviours are different .. ... to assert that an ANN that read barcodes was consciously experiencing ... implausible the humans are simply biological mechanisms following the 'laws' ...
    (talk.origins)
  • Re: Ben G on reinforcement-learning and the wirehead problem
    ... "Reinforcement learning is learning what to do--how to map ... It is of quite unnecessary for the agent to have any opinions about, or evaluations of, any of its behaviours. ... It is merely necessary for the system's architecture to enable conditioning of behaviours. ...
    (comp.ai.philosophy)
  • Re: Finding useful functions- part 1
    ... learning is related to the changes ... I don't think that for the behaviorist unsupervised learning does not ... since much development requires external inputs. ... modifies the repertoire of behaviours that may be shaped, ...
    (sci.cognitive)
  • Re: Gradual Learning, not Reinforcement Learning
    ... You seem to believe that "learning" means acquiring behaviours the system has never exhibited before. ... You use learning HORSE as an example, and posit HOXEL as an intermediate stage, with HORSE as an "ideal" to be achieved. ... Firstly, the system must be capable of "producing language", else it cannot learn the semantic difference between HORSE and HOTEL, both of which are easy consequences of HOXEL. ...
    (comp.ai.philosophy)