perm filename CONTAI[4,ALS] blob
sn#050693 filedate 1973-06-27 generic text, type T, neo UTF8
00100 This number is to be used for the new 4-output tables which are to
00200 be gated during learning in accordance with my new ideas.
00300 The general idea is to have about 16 large tables, ' am settling on
00400 256 cells at the present but every thing will be done to make it easy
00500 to change this number later if desired. There will be actually three
00600 sets of tables, one set used during recognition only, a second set
00700 used during learning only and a reserve set which is used to consolidate
00800 new learning with old at the end of each learning session or part of
00900 session and the recognition set updated
01000
01100 There will then be three distinct phases,- a learning phase, an update phase
01200 and a recognition phase. Each table will have each word divided into
01300 4 fields corresponding to the 4 outputs. In the learning tables these 4
01400 fields will be used to accumulate counts so that one can learn only
01500 until one counter has reached 511. Actually the counters may
01600 start with 31 rather than 0 so that the effective capacity is reduced
01700 to 480. This comes about from the fact that the reserve tables are used
01800 to count in units of 32 and after each update the overflow mod 32 is
01900 restored to the learning tables. The total capacity of the reserve and
02000 learn tables to accumulate counts is thus 2↑14, allowing for
02100 the 4 bit overlap. The recognition tables
02200 also have 9 bits in each field and so are used to contain probabilities
02300 in parts per 512.
02400
02500 Each table (except for the first one) has a parent table and a parent
02600 output field in this table. During recognition, the probability from
02700 this parent field is used to multiply the probabilities read from
02800 the line specified by the input and these probabilities are reported as
02900 the outputs from the table in question. In this way the final probabilities
03000 are kept comparable and reflect the true probabilities of the output in
03100 question.
03200
03300 If we limit the size to 256 as proposed, we could then allow for 4 inputs
03400 of 2 bits each, for 3 inputs of 3,3 and 2 bits, or 2 inputs of 4 bits each.
03500 It would be desirable to allow one to specify the choice simply by the
03600 number of input names that one typed in to the MAKE program.
03700
03800 As a further simplification to the MAKE program we could use the same names
03900 for table outputs as are used for phonettes anf features and arrange for
04000 the creation of the necessary linkages when one typed one name as that of
04100 the parent.
04200
04300 To simplify the use of different subsets of the tables, the header information
04400 would be kept separate from the tables proper. This information would then
04500 be read in but once and it would not have to be written out at the end
04600 of a run.
04700
04800 The handling of input tables if they to be used would be done by a separate
04900 MAKE program.