M.B. Clowes & R.W. Ditchburn (1959)
An Improved Apparatus for Producing a Stabilized Retinal Image, Optica Acta: International Journal of Optics, 6:3, pp 252-265, Taylor & Francis, DOI: 10.1080/713826291 http://dx.doi.org/10.1080/713826291

Abstract:
Criteria for defining the efficiency of an apparatus for stabilizing the retinal image are formulated. A distinction is made between geometrical stabilization and stabilization of illumination. A new technique is described which employs a telescopic normal incidence system. This makes it possible to obtain geometrical compensation both for rotations and for translations of the eye. It also gives good illumination stabilization. The degree of compensation achieved may be evaluated by precise physical measurements. About 99.7 per cent of natural eye rotations in horizontal and vertical planes is compensated and the effect of translations is negligible. The apparatus is designed to permit easy interchange of normal and stabilized viewing conditions.

NOTES:
(a) The Acknowledgments section states:

I don't have access to the next two papers, apparently written at NPL and referenced in his 1967 paper, below.

• Clowes, M. B., & Parks, J. R.(1961) A new technique in character recognition. Computing Journal , 4, 121-126.

• Clowes, M.B.(1962). The use of multiple auto-correlation in character recognition. Optical Character Recognition. Fischer, Pollack, Radack & Stevens, Eds. Pp. 305-318. Baltimore: Spartan Books, Inc.

AISB is still flourishing http://www.aisb.org.uk/ and celebrated its 50th year in 2014: www.aisb.org.uk/events/aisb14 .

Note: AI did not start at Sussex until Max arrived around 1969, and it did not start at Essex until Pat Hayes and Mike Brady, and later Yorick Wilks, established it around 1972 and after. So initially the "leadership" must have involved Edinburgh plus Max Clowes, then at Oxford? PDF versions of the conference proceedings (from 1974) are available at
http://www.aisb.org.uk/asibpublications/convention-proceedings

For information about the Machine Intelligence series see http://www.doc.ic.ac.uk/~shm/MI/mi.html

Max's address is given as M.R.C. Psycho-Linguistics Research Unit, University of Oxford The ACKNOWLEDGMENTS section states:

He discusses the work done in linguistics on formally characterising linguistic structures (e.g. spoken or written sentences) at different levels of abstraction, and remarks:

(... or selecting labels from a fixed set, he might have added).

Max's comment remains relevant to a great deal of 21st Century AI research in machine vision (i.e. up to 2014 at least), focusing on training machines to attach labels as opposed to understanding structures (I would now add "and processes involving interacting structures"). One of the points he could have made but nowhere seems to have made, is that natural vision systems are mostly concerned with motion and change, including change of shape, and change of viewpoint. The emphasis on static scenes and images may therefore conceal major problems e.g. those pointed out by J.J.Gibson. AI vision researchers later started to address this (partly influenced by Gibson), though as far as I can tell, it never interested Max.

So he specifies the objective of

In particular Max (like several AI vision researchers in that decade) argued that images had a type of syntax and what they depicted could be regarded as semantic content. Max attempted to develop a research methodology inspired partly by Chomsky's work, emphasising the importance of concepts of
- 'ambiguity' (two possible semantic interpretations for one syntactic form)
- 'paraphrase' (two syntactic forms with the same semantic content),
- 'anomaly' (syntactically well formed images depicting impossible semantic contents -- impossible objects, e.g. "The devil's pitchfork" often misdescribed as an "illusion"!)

He also emphasised important differences between pictures and human languages, e.g.

In section 4.5 he writes "we are characterising our intuitions about picture structure, not erecting some arbitrary picture calculus." This leads to the notion that the same picture, or portion of a picture, may instantiate different qualitative, relational, structures at the same time, i.e. different 'views'.

Unlike the majority(?) of current computer vision researchers he did not simply accept the properties and relationships that are derivable via standard mathematical techniques from image sensor values and their 2-D array co-ordinates (e.g. defining "straightness" in terms of relationships between coordinates in a digitised image) but instead attempted to identify the properties and relationships that are perceived (consciously or unconsciously) by human viewers and used in interpreting visual contents, i.e. working out what is depicted (the semantics).

This approach has important consequences:

In view of his more explicit discussions in later papers, I take him to be saying here that we expect to see things that are not in the picture but are represented by the contents of the picture. That would include, for example, seeing 3-D structures or object-fragments represented in a picture: the plane surface in which the picture lies can include only 2-D entities and their 2-D relationships.

The 1971 paper (listed below) is unambiguous on this point: the entities represented in the picture (the picture's "semantic" content) have 3-D structure, namely polyhedra, whose surfaces lie in different planes, most of which are not parallel to the picture plane.

What sorts of entities a collection of lines is intended to denote can affect how it should be parsed. E.g. he points out that in a circuit diagram, straightness of lines, and the existence of corners are less important than they might be in other pictures (e.g. a drawing of a building).

I am not sure whether Max drew the conclusion that instead of totally general purpose learning mechanisms applied to the raw data of visual and other sensors, human-like intelligent machines would need to have learning mechanisms tailored to the kinds of environments in which we evolved, and preferences for types of "syntactic" and "semantic" ontologies that have been found useful in our evolutionary history. Research on learning using artificial neural nets may be thought to meet that requirement, but that could be based on misunderstandings of functions and mechanisms of biological brains. Compare John McCarthy on "The well designed child".

At that time, I don't think Max knew how much he was echoing the viewpoint of Immanuel Kant in The Critique of Pure Reason (1781). However, he was aware of the work of von Helmholtz (perception is "unconscious inference") and he may have been aware of M.L.J Abercrombie's influential little book Abercrombie (1960), which made several similar points from the viewpoint of someone teaching trainee zoologists and doctors to see unfamiliar structures, e.g. physiological fragments viewed in a microscope.

Max later acknowledged the connection between his work and Kant's philosophy in Footnote 2 of the 1971 paper 'On Seeing Things' (listed below).

The Acknowledgments section of the Machine Intelligence 4 paper states:

In Oxford, Max had worked with Stuart Sutherland, who later came to Sussex University as head of the Experimental Psychology (EP) laboratory in the School of Biological Sciences (BIOLS). This functioned more or less independently of the social, developmental, and clinical psychology groups in schools within the Arts and Social Sciences "half" of the University.

A result of Sutherland's arrival was that Max was invited to return to the UK to a readership in EP, where he arrived in 1969. Somehow I came to know him and he, Keith Oatley and I had a series of meetings in which we attempted to draft a manifesto for a new multi-disciplinary research paradigm, including AI, psychology, philosophy and linguistics.

Robin's work shifted from AI to more "central" computing science thereafter.

M.B. Clowes, On seeing things, Artificial Intelligence, 2, 1, 1971, pp. 79--116, http://dx.doi.org/10.1016/0004-3702(71)90005-1

This developed the themes summarised above and introduced the line-labelling scheme used in interpretation of pictures of polyhedra, independently discovered by David Huffman Huffman (1971) , and referred to by Clowes in Footnote 1, which states:

This shared idea is often referred to as "Huffman-Clowes" labelling, and was generalised by many later researchers, including David Waltz who enriched the ontology and showed that constraint propagation could often eliminate the need for expensive search, and Geoffrey Hinton who showed that use of probabilities and relaxation instead of true/false assignments and rigid constraints, allowed plausible interpretations to be found in the presence of noise (e.g. missing or spurious line fragments or junctions) that would not be found by the alternative more 'rigid' mechanisms.

One of the themes of the paper reiterates the syntax/semantics distinction made in his earlier papers, emphasising the need for different domains to be related by the visual system, e.g. the picture domain and the scene domain, also referred to as the 'expressive' and 'abstract' domains. Consistency requirements in the scene (abstract) domain constrain the interpretation of the previously found structures in the picture (expressive) domain. An example in the paper is that in a polyhedral scene an edge is either concave or convex but cannot be convex along part of its length and concave elsewhere when there is no intervening edge junction.

The paper echoes his earlier paper in claiming that the interpretation of complex pictures requires "a parsing operation on the results of context-free interpretation of picture fragments" i.e. picture elements and their relationships need to be described, as a basis for interpreting the picture.

[This not a comprehensive summary of the contents of the 1971 paper.]

[ ... summary to be expanded ... ]

The Acknowledgments section thanks R. Stanton, A. Sloman and especially Jack Lang, "for exposing deficiencies in the formulation by attempting to program earlier versions of the algorithm".

NOTE on Hippolyte Taine

The paper had an important relationship to Max's ideas. I had long been interested in the role of diagrams in mathematical reasoning, especially in Euclidean geometry, and also topology, logic and arithmetic, and my 1962 DPhil thesis was an attempt to defend Kant's view that important kinds of mathematical reasoning could produce knowledge that was not empirical yet not just a matter of definitions and their logical consequences.

Max had observed that some of the details of human perception of pictures could be inferred from what we found ambiguous, synonymous, or anomalous, as explained above in connection with Chomsky's influence on his ideas. A visual "anomaly" occurs when 2D pictures have parts that are capable of representing 3D objects while the whole picture is incapable of doing so, just as some phrases or sentences have parts with normal semantic content whereas the whole phrase or sentence cannot, e.g. "John is (entirely) in the kitchen and (entirely) outside the kitchen". Pictorial examples include the Penrose triangle, the "devil's pitchfork" and various others. The connection between inferences and contradictions is well known in the case of sentences: e.g. if the joint truth of A and B is incompatible with the truth of C then A and B together imply not-C, and vice versa. This idea can be extended to contents of pictures. I don't know if Max made that connection, though several AI researchers have, and have studied diagrammatic reasoning as an alternative to logical or algebraic deduction, e.g. as reported in Glasgow et al. (eds) 1995, and elsewhere.

The image on the left depicts a possible 3-D scene.
Modifying it as on the right produces a picture that, if interpreted
using the same semantic principles, represents an impossible 3-D scene
(where blocks A, B, C form a horizontal line, blocks D, E, F form a vertical line,
G and H are between and on the same level as A and D, and the new block X
is both co-linear with A, B, and C, and also with D, E, and F).
The drawing on the right was by Swedish artist, Oscar Reutersvard, in 1934
http://im-possible.info/english/articles/triangle/triangle.html

So a complex picture made of parts representing possible 3-D configurations may have N parts such that if a certain part X is added (e.g. an extra line joining two of the junctions, or a picture of an extra block that is simultaneously co-linear with two other linear groups, as in the above figure), then it becomes anomalous and cannot represent a 3-D configuration using the same rules of interpretation (based roughly on reversing projections from 3-D to 2-D). In other words the original N parts have a joint interpretation that entails that the situation depicted by adding the part X cannot exist. This is analogous to logical reasoning where N consistent propositions entail that an additional proposition X is false. So its negation can be inferred to be true. This example could not be handled by the Huffman-Clowes system, as it requires a richer grasp of geometry. Humans can reason that the configuration on the right is impossible without knowing any of the actualdistances or sizes, whereas I don't believe any current AI vision system can do that. (This is one of very many forms of geometrical and topological reasoning that are still beyond the scope of AI systems. Moreover I don't think neuroscientists have any idea how brains can support this kind of reasoning.)

As far as I know, Max provided no explanation of how such impossiblities (anomalies) are recognized by humans. What cognitive mechanisms and processes underly the intuitions on which the linguistic and the geometric examples depend remains an unanswered question, though automated theorem provers can deal with logical and arithematical inferences. The "line labelling" algorithm presented in Clowes (1971), which had been independently discovered by David Huffman, enabled a computer to derive the impossibility of certain 3-D interpretations of picture fragments from the fact that implied combinations of edge and junction features that were ruled out by programmer supplied rules. Those rules were discovered by humans thinking about 3D volumes bounded by planar surfaces. As yet I don't know of any AI program that can discover such geometrical constraints itself, though there may be some mathematically equivalent arithmetical theorem about coordinates of collinear or co-planar sets of points that a machine could prove. Note that training a program on a huge collection of pictures in which the anomalous cases do not occur might give the program the ability to assign a low probability to such cases, but would not give it a human-like understanding that they are impossible, and why they are.

There have been extensions of the Huffman-Clowes idea to reasoning about curved objects, and scenes involving more complex polyhedra along with cracks and shadows, e.g. Waltz.

Moreover the ideas were extended by Waltz, Hinton, Mackworth and others to allow constraint-propagation techniques to be used to improve efficiency, and to allow use of 'soft' constraints so that missing or extra features caused by poor lighting or use of low quality cameras could allow most of the image to be interpreted correctly despite missing or spurious fragments. In all these cases human reasoning was used to discover the rules to be used, rather than reasoning by machines with understanding of spatial structures.

My hope in 1971 was that we would one day understand how to build a 'baby' robot with abilities to interact with its environment and with a suitable set of innate mechanisms for extending its knowledge and modes of reasoning in something like the way human children and mathematicians do. This could explain why Kant was right about the nature of mathematical knowledge, by demonstrating a working system, able to use spatial reasoning to make a variety of deep mathematical discoveries, including a significant subset of topology, Euclidean geometry and arithmetic. I also thought that in some contexts such non-logical forms of reasoning would not only be closer to human mathematical reasoning than logical deductions from axioms, but might add additional power to intelligent machines, in particular heuristic power based on search spaces closer to the problem domain. I don't know whether Max agreed with this, but it was he who pressed me in 1971 to submit the paper criticising purely logicist AI that was accepted for IJCAI.

Since writing the paper I have discovered that it is very difficult to get computers to emulate the forms of spatial reasoning that are common in human mathematics, although some special cases can be handled by translating geometry into arithmetic and algebra following Descartes and using modern logic. But that's not the method used by Euclid centuries before Descartes, Frege and modern logicians. (For anyone interested, some topological examples are discussed here.) There are also no working models capable of explaining or replicating the spatial reasoning abilities of nest building birds (e.g. weaver birds), elephants, human toddlers, and many other intelligent animals.

A talk with this title was given at the Institute of Contemporary Arts (ICA), as part of a series of lectures at the ICA in 1971-2. Papers corresponding to the lectures were collected in the book edited by Jonathan Benthall (which also included a paper on AI language processing by Terry Winograd). Although the phrase "controlled hallucination" is not in Max's paper, the idea is there.

The talk included a still from John Schlesinger's film "Sunday bloody Sunday" in which two intertwined bodies were shown, presenting the viewer with the task of using world knowledge to decide which parts belonged to the same body: an example of creative "controlled hallucination" of hidden connections. The same picture is in the published paper. My amateurish sketch of the scene is below.

Which hands belong to whom?
Answering this question requires use of knowledge
of human anatomy to hallucinate unseen connections.

However I acquired the book in April 2014, and searched through the paper. There is no mention of hallucination, though Max may have used the phrase "controlled hallucination" when presenting the talk at ICA. Readers should be able to use their own controlled creativity to hallucinate invisible connections in order to work out which hands belong to which person, using knowledge of human anatomy.

That conference paper was later revised for a journal:
Frank O'Gorman, M. B. Clowes: Finding Picture Edges Through Collinearity of Feature Points. IEEE Trans. Computers 25(4): 449-456 (1976) https://www.researchgate.net/publication/3046601_Finding_Picture_Edges_Through_Collinearity_of_Feature_Points

Frank O'Gorman: Edge Detection Using Walsh Functions. Artif. Intell. 10(2): 215-223 (1978)

Steve Draper followed up a BSc in Physics and a DPhil in AI at Sussex supervised by Max, with a career in psychology. A paper based on his DPhil work (funded by one of Max's projects) was published in 1980:

Larry Paul was supervised by both Max and Aaron Sloman.

Frank O'Gorman, after an MSc in Computer Science at Birmingham, worked with Max as a research assistant for several years (and also taught me a great deal about programming and computer science, partly by teaching me Algol 68, which was used for their research in the mid 1970s). After the grant ended he worked for a while on my POPEYE project described here, which attempted to extend some of Max's ideas. The grant was awarded by the Cognitive Science panel of the Science Research Council on condition that Max had an advisory role. Others on the team were David Owen, and, for a short time, Geoffrey Hinton. We were all deeply influenced by Max. [Other students, collaborators, etc. ... ?]

• The Sussex University Bulletin obituary notice in 1981: http://www.sussex.ac.uk/internal/bulletin/downloads/1980-1989/1981/May/19810512.pdf mentions that Max was at Reading, Oxford, and NPL, before he went to Australia.
• There was a short notice in the proceedings of IJCAI 1981, for which I do not have the text.
• Experiencing Computation: A Tribute to Max Clowes, by Aaron Sloman, was published in Computing in Schools in 1981, the same year as the BBC Micro was announced.
• The longer version included above was published in 1984, in New horizons in educational computing, Ed. Masoud Yazdani, Ellis Horwood Series In Artificial Intelligence, pp. 207--219, Chichester, http://www.cs.bham.ac.uk/research/projects/cogaff/00-02.html#71