perm filename MIT[901,BGB] blob
sn#129628 filedate 1974-11-12 generic text, type T, neo UTF8
00100
00200 NOTES ON A VISIT TO M.I.T 3 - 4 - 5 JULY 1969
00300
00400 B.G.Baumgart
00500
00600 I was most impressed by Richard Greenblat. Although he is
00700 basically inarticulate - he is excellent at demonstrating
00800 things in great detail - from midnight to four he demonstrated
00900 and mumbled about the following:
01000
01100 A plausibly good edge finder based on taking gradients of the
01200 intensities of a whole image. Greenblat apparently developed
01300 the edge finder by getting a feel for the data and partial
01400 results using good diagonostic display on the DEC 340. For
01500 example one display showed the input image in eight intensity
01600 levels over which he could run a box which selected an area
01700 for detailed examination of the gradient vector field within
01800 the box - A second display consisted of a line segment
01900 which he could translate and rotate over the image in order to
02000 select a domain for his local-edge-detecting function which
02100 was monitored by two cartesian graphs and a list of eight quantities
02200 of merit to be maximized, minimized or set to good values - e.g
02300 averages, maxima, ratio of maxima to neighbors, area under curve from
02400 half maxima thru maxima to half maxima, etc.
02500
02600 Greenblat apparently has contrived an edge operator by selecting or
02700 rather by tuning the values of various cheat-factors, bias-constants,
02800 fudge-variables and petty heuristics (if such&such then forget about
02900 computing mumble). The operator selects likely edge-points with no
03000 concept of straightness or corners - Another operator goes around
03100 fitting the edge-points into lines. And Greenblat was going to do
03200 corners, etc. ,but he had to first make some diagonostic displays.
03300
03400 "Are you attempting to creat input for Guzman ? " , I asked.
03500 "Yes, but I will rewrite Guzman in machine code.", he replied.
03600
03700 Richard further demonstrated for me many features of operating the
03800 two image dissectors: Focusing, Zooming (with the zoom lens and with
03900 the dissectors internal resolution) as well as much trivia.
04000 Then he demonstrated and explained chess & its debugging displays -
04100 Greenblat has no new version to export at the moment. Chess is probably
04200 a prime love of his, attempts to change the subject were ignored in the present
04300 but turned out to have been accurately and thoroughly queued.
04400
04500 So there were displays of Chess Alpha-Beta Stuff, Evaluations, etc.
04600 Greenblat would say: " power is inversely weighted queens low
04700 pawns high in this place here (he points to a dense flickering CRT)
04800 ...and here (he points to a listing that looks like 7-bit ASCII printed
04900 on a Flexowriter) " His technical statements would be terminated by
05000 a "you know ? " and an expectant look. I would answer mostly positively
05100 with a nod, smile, "yes", "a-ha", etc ... I would occassionally
05200 answer negatively and generate a question "No I don't understand
05300 the Queen's Mummble - tell me about it."
05400
05500 Later we demonstrated the Color Display - it works but isn't being used
05600 - point mode rather than a vidicon raster scan; Music - plays from
05700 core in real-time has good translation package of middle-C-sharp
05800 notation to cycles per second with harmonics notation; Microlights -
05900 the arm has little lights for visual tracking - no code yet; as well
06000 as the MIT model railroad telephone exchange controller - "See the
06100 train is coming to switch 17, so I punch 17 here and ..."
06200
06300 Richard has worked gung-ho full time for the past two weeks on
06400 doing the TOWER-COPY-DEMO which he thinks will be a plateau or
06500 temporary stopping place for Robot Research at M.I.T. - since he
06600 says A.I. types would really rather do other things such as chess
06700 or natural languages or mathematical A.I. theory. Richard will
06800 go back to chess of which he is quite fond and still has many ideas
06900 for further improvements. The color-hydraulic slide box
07000 doesn't work and isn't being used, but it is considered interesting and
07100 is not available for transcontinental loan just yet.
07200
07300 I believe that Greenblat is an exemplar AI worker who can be followed.
07400 His method and attitude include:
07500 1. SHUT-UP-AND-DEAL-WITH-THE-CODE ism
07600 2. SUPER-DISPLAY-DIAGONOSTICS - take an incremental
07700 step forward and then see what things look like from
07800 there.
07900 3. INCLUSION-PLAGARISM attitude - most AI types seem to have
08000 one good idea (if any) with which they try to conquer all
08100 parts of all problems - Greenblat will take any and all
08200 ideas that are around and use them. This passes over rather
08300 than solves the problem of getting people to Work together.
08400
08500 A final consideration is what sacrifices have to be made in order to
08600 be a greenblat-programmer, I will relate my subjective thoughts
08700 on this to any would-be's.
08800
08900 GOSPER can demonstrate picking up cubes. He's against
09000 "human-wave-robot-research", likes Time-variant displays, thinks
09100 Time-Sharing & Robots are related but that it is unnecessarily difficult
09200 to research & develope both together. He anticipates getting one processor
09300 dedicated to the Arms while the other does time-sharing LISP & editing.
09400 He doesn't think Guzman's program is useful - no one can generate input
09500 to Guzman and he can't see how the output would be useful - Minsky
09600 however is working on a paper about Guzman pathology (or something) titled
09700 "See" - Minsky likes Guzman's because it does so well and knows so little,
09800 Gosper thinks Minsky is too partial towards elegant mathematical solutions.
09900
10000 Gosper believes what we really need to work on is Cerebellum
10100 functions - such as physical coordination & visual lobes processing
10200 rather than Cerebrium mathematics and intellect. That
10300 our programming instincts are misleading and wrong in that we think in
10400 terms of minimal, clean logical solutions rather than processing thru
10500 redundant data and then forming a conclusion when it is demanded
10600 ready or not.
10700
10800 Gosper, admires Nick Horn's energy with respect to cataloging
10900 Robot diseases and real-world distortions.
11000
11100 Random Gosper Comments:
11200 i) Slowness of the image dissector is not screwing us - if we were
11300 really doing vision we would be compute bound anyway.
11400 ii) The image dissector is not likely to be damaged by excessive light,
11500 the so-called 29-craters(non-responsive pinpoints) were
11600 probably there at delivery and were caused by particles of dust.
11700 He is said to have a kludgy line follower and to be doing incremental
11800 work on block stacking.
11900
12000 DAVID SILVER is a real live intelligent high school dropout and radio
12100 hacker: i) he has done a computer controlled mouse that herds balls of
12200 paper around the floor using vision to see "Blobs" of paper. ii) He
12300 is working on a Garbage-Collector which picks up wads of paper, puts
12400 them in a waste basket, dumps the basket out and then starts over. iii) He
12500 is considering motion by pointwise subtraction iv) He interfaced the Color TV to the
12600 PDP-10, The color TV works I saw spacewar and a number of demo displays -
12700 color mixing is negligible, display points don't allign properly.
12800
12900 WITT DIFFIE Math - Lab Lisp guru, has made drastic revision to the LISP
13000 compiler (a copy of which I have on a DEC tape) wants to leave MIT for
13100 STANFORD, I told him to call John McM. or Les.
13200
13300 HOLLOWAY is working on the over head arm system. HOLLOWAY, LEVITT &
13400 EASTLAKE are in a small, company Systems-Concepts-Inc. which is marketing
13500 multi-terminal-graphics and is trying to move to San Francisco.
13600
13700 TOM BINFORD is working on a perspective line follower, regions, local
13800 homogeniety (similar points are grouped together) edges - theory.
13900
14000 FREDKIND - Triple I administrator who tried to reorganize and to
14100 discipline AI work at MIT and isn't seen around much anymore.
14200
14300 KRAKAUER is studying intensity contouring of curve surfaces such
14400 as fruit - with a single light source the intensities are a clue to
14500 depth perception. I saw recognizible cal-comp contour maps of
14600 a pear - this requires the intensity precision of which the image dissector
14700 is capible.
14800
14900 NICK HORN has a focusing routine that does local focusing for the sake
15000 of depth perception - using a fourier transformation (Greenblat says
15100 he will use it) - the idea is solidly based on considerible analysis which starts as follows:
15200 a point on the object reflects light which is gathered by the lens and focused
15300 to a point on an in-focus image plane - now if the image is out of focus
15400 the light for a single object point form a circle and the unfocused image
15500 is the summation of such circles - in order to focus the
15600 image you transform, attenuate (multiplication), and then transform. I
15700 have the paper on this and will try some of it and would appreciate
15800 help ( At the ARPA junior achievement conference I learned on the bus
15900 that at Utah transform-attenuation-transform is used for enhancing x-ray
16000 pictures ).
16100
16200 DAVID WATTS demonstrated a pressure sensitive hand - deformations
16300 of a soft coaxial cable caused reflections of a pulse - a program
16400 could display pressure and locus.
16500
16600 RICHARD ORB is a summer worker attempting to eliminate shadows from
16700 Guzman data from characteristics of vertices & intensites of regions
16800 local to a Guzman Anamoly or around inconsistent places in the input.
16900
17000 ADOLF GUZMAN complained that he couldn't get his drawings inputted - and
17100 considers his problem to be to get some real input.
17200
17300 MARTIN - math lab, not very talkative.
17400
17500 EUGENE CHARNIAK calculus word problems & story abstracting.
17600
17700 TOM JONES electric baby strategy - AI stagnation work - goals,subgoals,
17800 5 and 10 cent psychology, plausible moves, etc.
17900
18000 CARL HEWITT - PLANNER: A language for manipulating models and proving
18100 theorems in a Robot - A man and a paper in LISP unbound to reality.
18200
18300 MINSKY is working on Guzman anamolies & perceptrons.
18400
18500 PAPERT was observed leaving his office - he is known for being
18600 hard to reach and no one was too sure about what he was doing.
18700
18800
18900
19000 I have a listing and a DEC tape of MIT's Incompatible Time Sharing
19100 as well as afew selected MIT-AI-Memos.
19200
19300