Transcript document

Cognitive Computing
2012
The computer and the mind
6. DREYFUS AND DREYFUS
Prof Mark Bishop
Overview

The paper opens with quote from Wittgenstein:

“Nothing seems more possible to me than that people someday will come to the
definite opinion that there is no copy in the ... nervous system which corresponds to a
particular thought, or a particular idea, or memory.”


Ludwig Wittgenstein, 1948: last writings on the Philosophy of Psychology,
Volume 1.
And from Rumelhart (& Norman) on distributed representation:

“Information is not stored anywhere in particular. Rather it is stored everywhere.
Information is better thought of as ‘evoked’ than found.”


08/04/2016
Rumelhart & Norman, (1981).
Albeit that such a ‘distributed representation’ remains spatially located.
(c) Bishop: The computer and the mind
2
Symbols and Connections

In the late 1940s and early 1950s, people realised computers can do more
than just perform arithmetic.


In 1950 Turing published his seminal paper ‘Computing Machinery & Intelligence’ on Artificial
Intelligence (AI) which took seriously the notion that one day ‘we would speak of machines
thinking without fear of contradiction’.
From the 1950s two schools of (Artificial) Intelligence (AI) emerged:

Symbolic intelligence:
 Easiest to instantiate a formal world model;
 Philosophically rationalist & reductionist.

Connectionist intelligence:
 Easiest to model the mind (as interactions of neurons);
 Philosophy holistic; linked to cognitive neuroscience.
08/04/2016
(c) Bishop: The computer and the mind
3
Symbols and intelligence

Symbolists view minds & computers as
physical-symbol systems


Newell & Simon’s hypothesis
This idea is philosophically grounded upon
work of Frege, Russell & (early) Whitehead

08/04/2016
Which are in turn heir to the ‘atomistic
reductionist’ tradition in philosophy.
(c) Bishop: The computer and the mind
4
Atomism and reductionism:
Descartes & Hobbes

From Descartes:



Understanding as the forming and manipulation of
appropriate ‘mental representations’ formed from
primitive elements,
 ‘naturas simplices’ - ‘simple elements’
And all phenomena are understood as combinations of
these.
From Hobbes:


08/04/2016
We deduce that these ‘simple elements’ were formal and
related by purely syntactic - rule based – operations;
Hence all reasoning is reducible to calculation;
ratiocination.
(c) Bishop: The computer and the mind
5
The symbolist A.I. research
programme

From Leibniz


To avoid an infinite regress there must be at base simple elements which
represent things in the world, and which are mixed together to define
complex objects – an alphabet of human thought.
From (early) Wittgenstein

The picture theory of meaning (atomistic; as defined in the ‘Tractatus’)





“World is the totality of facts, not of things”;
Such ‘facts’ are logically described via (syntactic) pictures;
Elements of such pictures being combined in definite ways to represent the
ways things are combined;
The ‘Tractatus’ was seen as the culmination of the rationalist tradition.
Symbolic AI is the attempt to find such primitive elements;

08/04/2016
Newell & Simons hypothesis effectively turns Wittgenstein’s early vision
into an empirical claim and bases a research programme on it.
(c) Bishop: The computer and the mind
6
The opposing tradition …

Took inspiration from neuroscience not philosophy ...


E.g. Hebb’s early work on machine learning; Hebbian learning.
… And Rosenblatt, who conceived that it would be
easier to formalise the brain and then investigate its
behaviour, rather than attempt to formalise behaviour
and then design an axiomatic system to implement it.
08/04/2016
(c) Bishop: The computer and the mind
7
A declaration of ‘war’ !

It is possible to consider the symbolists as people
building ‘problem solving machines’.

Whereas connectionists wanted to build systems to
‘generate their own behaviour’.

Initially both systems appeared successful:


as early as 1958 Herbert Simon claimed he had machine that
‘can think’ ..
.. and by 1959 Rosenblatt publicised much the same opinion
of his machines.
08/04/2016
(c) Bishop: The computer and the mind
8
Problems for A.I.

Following the publication of Minsky and Papert's monograph
‘Perceptrons’, the symbolists seemed set to ‘win the war’ of A.I.

Yet both traditions had their detractors:

Connectionism via Minsky and Papert’s 1969 book, ‘Perceptrons’.

Symbolism via:



The 1973 UK Lighthill Report, ‘Artificial Intelligence: A General Survey’,
James Lighthill, Artificial Intelligence: a paper symposium, [UK] Science
Research Council.
Dreyfus, Hubert (1972), What Computers Can't Do, MIT Press.
And the detractors essentially much made the same point:

08/04/2016
That extant A.I. systems only work on TOY problems.
(c) Bishop: The computer and the mind
9
A philosophical crusade

Minsky and Papert's attack was seen by some as a philosophical
crusade:



M&P’s 1969 analysis was only of ‘Single Layer Perceptrons’ and yet it
succeeded in virtually stopping all research into connectionism:


reductionism was being challenged by ‘evil’ neural holism;
atomists need ‘hidden nodes’ to refer to symbolic [micro] features of
environment – connectionists are not so committed.
Rosenblatt’s neural research was discredited and connectionism was
not even mentioned in early edition of Margaret Boden’s seminal text
on AI, ‘AI and natural man’.
However there are other reasons for this prejudice:


08/04/2016
With only limited computing power early symbolists could do more;
There is a persistent belief that thinking and pattern recognition are
separate and that ‘thinking’ is more important.
(c) Bishop: The computer and the mind
10
‘Atomism’ in A.I.

The underlying philosophical idea - from Plato through Leibniz - is
that understanding a domain entails having a ‘context free’ theory
of the domain;

enabling [relatively] easy knowledge transfer from one domain to another.

Winograd famously described A.I. as an attempt to find a formalism
for knowledge, identify the atoms from which it is built and the
forces that act upon it.

At the time no one specifically argued for atomism in AI, there just
remains an implicit assumption that, because it works in other
domains it will work here.
08/04/2016
(c) Bishop: The computer and the mind
11
Problems for simple atomism /
reductionism

However the conclusions of later Wittgenstein …


After publishing ‘The Tractatus’, Wittgenstein spent several years doing ‘phenomenology’ looking for ‘base atoms of meaning’ - yet ended up abandoning rationalist philosophy
altogether.
And the work of early Heidegger …

For Heidegger traditional philosophy is defined from the start by its focusing on facts (in the
world) while "passing over" the world as such.



I.e. For Husserl, an act of consciousness – noesis - does not grasp onto the object itself, (which is famously left
‘bracketed’), but the noema - an ‘abstract form’ (effectively a hierarchical representation of ‘facts’ [of the world])
correlated with the act [of directed consciousness].
Whereas Heidegger reasoned that it was fundamentally impossible to find such context-free
elements of meaning - facts [of the world] - because depriving any element of its context –
passing over the world – deprives it of the very organisation that makes possible its [veridical]
use.
And this boded ill for simple reductionism!
08/04/2016
(c) Bishop: The computer and the mind
12
Husserl’s ‘Phenomenolgy’

For Husserl:

An act has ‘directedness’ only because of the intellectual
reasons that ‘give it meaning’;

Ones ‘predicate senses’ somehow pick out an object’s ‘atomic
properties’ [facts] which are subsequently hierarchically
combined to form complex descriptions of objects in the world;
08/04/2016

At the top level there is effectively a ‘rule’ defining all the features - and
properties - that can be possibly part of this type of object ..

.. a system analogous to Marvin Minsky‘s concept of ‘Frames’ – a
knowledge representation systems in classical AI.
(c) Bishop: The computer and the mind
13
Heidegger

In contrast to Husserl, Heidegger suggests there are other ways of
encountering things other than as ‘objects’ defined by a set of ‘predicates’
(context free ‘facts’)


For Heidegger in [trouble free] use; ‘smooth coping’ …



A mode of engagement Heidegger calls the ‘ready-to-hand’;
When things go awry (e.g. the nail breaks; the wall is too solid and requires a heavier
hammer) Heidegger refines the mode-of-engagement to the ‘un-ready-to-hand’).
… everyday objects – hammers; door knobs etc. - are defined by a context
of normative social roles (“the manifold assignments of ‘in-order-to’ ”)


A ‘mode of engagement’ Heidegger calls the ‘present to hand’.
I.e. In the everyday ‘ready-to-hand’ use of a hammer we actualise a skill with no clear
division of subject and object (represented in the mind) in the context of a socially
organized nexus of equipment, purposes, and human roles.
… and the “ ‘sight’ with which they accommodate themselves is
‘circumscription’ ”, (Heidegger).
08/04/2016
(c) Bishop: The computer and the mind
14
A context free theory of the world
?

Can there be a context free theory of the world or is the common-sense background rather an
‘impenetrable’ ensemble of skills, practises and judgements which cannot be explained in terms of
rules?

Husserl sought to answer this by asserting that ‘the background’ is just the interaction of millions
upon millions of rule-based axiomatic beliefs (which have truth conditions; and are ‘facts’ if true):



However by the age of 75 Husserl concluded that ‘Phenomenology was an infinite task’ - as he
had to include more and more of a subject's common-sense understanding of the everyday world
to describe it


So we can characterise the world by ‘detachment from it’ and then enumerating all such beliefs;
So completing the reductionist, atomistic philosophical task began by Socrates…
and there is some evidence Minsky felt similar wrt his ‘frame’ knowledge representation.
The naiveté of A.I workers regarding (contemporary) philosophical research led Hubert
Dreyfus to predict trouble for A.I. (Dreyfus, 1972, “What computers can’t do”); a book [until
recently] generally ignored by AI community.
08/04/2016
(c) Bishop: The computer and the mind
15
Three stages of (symbolic) AI

1. Representation and search;


2. Facts and rules;


but toy ‘micro-worlds’ did not prove scalable.
3. Common sense knowledge:


Via, for example, ‘means-end’ analysis.
A.I. believes common-sense is formalisable; but need it be?

Perhaps ‘common-sense’ is nothing more than a vast set of ‘special cases’ ?

Is even a ‘naive physics’ formalisable – or does a child perhaps simply learn to discriminate (deploy
‘neural judgement’) a large set of special cases?
Dreyfuys asserts that, “the rationalist tradition has been put to an empirical
test and failed”

Justifying Rosenblatt’s ‘connectionist’ approach (re. his intuition that rationalism would be
difficult);

Now even Terry Winograd has ‘lost faith’ with A.I. and teaches ‘continental’ philosophy..
08/04/2016
(c) Bishop: The computer and the mind
16
The new connectionism

Frustrated A.I. workers flock to connectionism:
 E.g. My own experience at the Oxford Experimental Psychology
conference.

NB. If the connectionists are correct then philosophers will have to
give up the atomistic, logicist, rationalist tradition.

Albeit neural-net researchers, influenced by symbolists, have tried
to find features of reality in hidden nodes; but this is only true in a
trivial sense (by assigning an invented name):
 Uni-variate neural codes; compare with SDPs bi-variate data.

Yet the connectionists only success is also only with limited models;
 Perhaps connectionism is simply getting a [deserved] chance to
fail, like symbolism…
08/04/2016
(c) Bishop: The computer and the mind
17
Common sense and connectionism

The ‘common sense knowledge problem’ in neural computing is that of
generalisation…



But what counts as a successful generalisation?

The neural-net designer has in mind what is a successful generalisation, but how is this
defined?

If a classifier network produces an output of an unexpected type, it might have merely
‘learnt’ a different definition of type to that the designer intended:
Ideally in generalising the network can either interpolate between points or
extrapolate beyond them;



for a neural net to be useful it must be able to ‘generalise’…
in reality an infinite number of curves can go in between and beyond such points;
Cf. intelligence tests.
Hence in engineering applications the neural network designer determines
an architecture to restrict possible transformations, such that net behaves
appropriately for the application in mind.
08/04/2016
(c) Bishop: The computer and the mind
18
Conclusion

If early Heidegger and later Wittgenstein are
correct then intelligence is much more holistic
(and social) than either connectionism or
symbolism imply.

Like symbolism, perhaps connectionist
systems need to be properly embedded in a
social reality to have any chance of making
intelligent progress…
08/04/2016
(c) Bishop: The computer and the mind
19