NetNews Usenet Archive 1992 #16/NN_1992_16.iso/spool/comp/ai/neuraln/2975

Internet Message Format | 1992-07-27 | 7.5 KB

Path: sparky!uunet!mcsun!uknet!warwick!str-ccsun!strath-cs!robert
From: [e-mail] (Robert B Lambert)
Newsgroups: comp.ai.neural-nets
Subject: Re: Correctness of NNs
Message-ID: <[e-mail]>
Date: 27 Jul 92 11:58:43 GMT
References: <[e-mail]> <arms.712022732@spedden>
Organization: Comp. Sci. Dept., Strathclyde Univ., Glasgow, Scotland.
Lines: 136

In article <arms.712022732@spedden> [e-mail] (Bill Armstrong) writes:
>[e-mail] (Robert B Lambert) writes:
>
>>Surely to be able to state that network X is 100% reliable, the entire input
>>set must be known. From experience, any pattern recognition task which has a
>>fully defined input set with appropriate responses can most easily be solved
>>with a look-up table.
>
>You are right, in theory, about table look up.  But try to do that for
>a 32*32 grid of pixels.  You would have to store 2^1024 values, which
>is not possible in this physical universe, and certainly is not
>economical.  This is typical of high-dimensional problems.
>
>So you can't store all outputs, you have to do some things by
>"generalization" of some kind, based on a stored state derived form
>some "training set", say.

Interesting argument. If you have an input grid with 32x32 binary pixels, then
indeed you have 2^1024 possible input patterns which is physically impossible
to implement by a look up table. However, say we are recognizing handwritten
characters and we have a database of 100,000 characters. We could use a look up
table (size 12Mbytes) which would give 100% correct classification for each
character in the database. If all future recognition is restricted to this set,
the look up table provides the simplest (and probably cheapest solution).

However, such a system is virtually useless if new characters are presented.
A better solution is a system capable of generalization. Such a system can be
implemented which gives 100% correct classification for the 100,000 characters
and gives a good classification rate for unseen characters. But what about
non-characters? Our original training set is a tiny percentage of all possible
input patterns. If we fire random patterns at the input, what will the network
do?

Any neural network which can only classify inputs into a number of pre-defined
categories will dramatically fail in the real world.

>>I had thought that the principal strength of neural networks (including the
>>brain) was the ability to form adaptive responses/rules based on a subset of
>>all possible system inputs. If responses are wrong, a neural network has the
>>ability to correct itself, whilst improving its response rate on subsequent
>>new inputs.
>
>Sounds ok.  But, between the lines I read that you think you could
>correct any errors on huge input spaces like the 2^1024  - sized one
>above.  That can't be done, in general, simply because you don't have
>enough memory to do it.  Certain functions you will never be able to
>compute in our universe.

True, including the correctness of most computer software. However if a network
is able to identify inputs which do not lie within expected ranges, appropriate
action can be taken. ie, if the input is not a numeric character, do not attempt
to categorize it.

>>With respect to safety in critical applications, how do you prove a system to
>>be correct? You first have to determine every possible input to that system.
>
>I agree, you have to specify a superset of the inputs you can get.
>
>>It is never possible to eliminate errors from any real system no matter how
>>good it looks on paper. The current approach to this problem is redundancy.
>>Build a number of systems from different component running different software
>>and make sure they all produce the same response during use. Is this not one
>>of the strengths of NNs? If a cell fails or a connection is broken, the
>>degradation of the response to each input is slight.
>
>Sounds great, but doesn't work.  There is a probability greater than
>zero in a large network that you will get very ungraceful behaviour.
>
>I think your statement about "slight" degradation is quite wrong.  You
>Should be able to cook up cases where one connection being wrong
>throws off a whole net.  For example, if an output unit has a weight
>that erroneously acts twice as large as it should be, you may not have
>such a graceful degradation.

Does this apply to the human brain? I think not. The human nervous system has
massive redundancy with the expectation of component and connection failures
over time.

Much of my arguments and reasoning is based on the study of the human brain.
ANNs are currently very primitive and limited in their application. The current
argument over analogue vs digital transmission of information is an excellent
example of the problem neural network researchers face. The individual neurons
in the brain are causal devices. Their state is based on the history of firing
of the cell and the history of received pulses. BP nets try to model the net
behaviour of each neuron by considering the effective output (ie output signal
proportional to input). ALNs as with many other model the binary outputs of
biological neurons.

It is to easy to get caught up with the workings of individual neurons and
forget the connectivity. ANNs (inc ALNs) do not model the biological brain.
There is strong evidence for cooperation between large number of neurons in the
brain utilizing very high levels of feed-back. Such `groupings' act together
to generate a global response to a given stimuli. For this cooperative grouping
to fail, a large percentage of the neurons within the group must fail.

If we want networks to both solve real world problems and improve our
understanding of the human brain, we should be considering the higher level
behaviour and connectivity of networks and not their basic components.

>But how can you claim that x-technology is able to produce
>sensible responses to new inputs as a general rule?  Show me a proof,
>say that fuzzy logic always produces sensible responses to new inputs,
>and I'll eat my hat.  First, there's a complete, all encompassing
>definition of "sensible", then....
>
>For those who are just getting into the discussion: please do not
>think that ALNs are safe but BP nets aren't.  Both can produce
>unexpected values not detected by testing.  I intend to show that ALNs
>can be used with a design methodology that will lead to safe systems.
>I will leave it up to the BP people to worry about the safety of their
>systems.  Up to now, it seems they won't even admit there is a
>problem.
>

I agree with you to a point. Simple NNs by there very nature are unreliable.
My comments about sensible responses do not refer to any single network, but
rather a principal we need to work towards. The human brain is capable of
identifying stimuli which do not conform to the norm. The action taken is based
on higher level reasoning and experience and can usually be defined as sensible.

What is the future of ANNs? If they are to be used in any situation where a
fully definable input-output set exists, they have no future as this is the
application where conventional computer technology excels. If ANNs are to be
used for real world control and recognition tasks we must face up to the fact
that such networks while able to give the best performance can never be 100%
reliable as it is simply not possible to account for all possible inputs.

If ANNs were used to replace drivers in cars and the number of fatalities were
reduced (but not removed) would this be acceptable?

-------------------------
Robert B Lambert
University of Strathclyde
Scotland, UK.

[e-mail]