Gary McGraw & Daniel Drasin. (1993) Recognition of Gridletters: Probing the Behavior of Three Competing Models. In Proceedings of the Fifth Midwest AI and Cognitive Science Conference, pages 63-67, April 1993.

Author of the summary: Patrawadee Prasangsit, 1999, pp@cc.gatech.edu

Cite this paper for:

• Letter Spirit: A project that models aspects of human high-level perception and creativity on a computer, focusing on the creative act of artistic letter-design.
• DumRec, NetRec, FnetRec - the three models being compared in the paper.

Summary:

Categorical sameness is property possessed by instances of a single letter in various styles (e.g., the letter 'a' in Times, Courier, Palatino).

Stylistic sameness is property possessed by instances of various letters in a single style (e.g., the letters 'a', 'b', and 'c' in Times).

Each letter is formed by a set of short line segments, called quanta, on a fixed grid of dimension 3x7.  See figure 2.

The three models are:

DumRec

• Associated with each training letter is a property list.
• Given a mystery letter, DumRec computes its property list and compares it with that of each training letter.  The score is weighted sum of the match of the property lists.
• The weights play a crucial role in DumRec's performance.  To modify them is to "tune" DumRec.

• 2- or 3-layer feedforward connectionist networks trained using backpropogation.
• 56 input units, each corresponds to a quanta.  26 output units, each corresponds to a letter of alphabet.  Hidden layer may have 0-120 units.
• Major open problems: a learning rate for backpropogation and the number of hidden units.

• A variation of NetRec.  Forces the network to pay more attention to certain features as determined by human.
• Train a number of small "subnets" to detect certain features.  Examples of features are height, weights, descenders, ascenders, different numbers of tips, etc.
• Input to the letter-recognizer network is the existing 56 input units as well as outputs from the subnets.

• The percentage of successful recognition: DumRec 74.3%, FnetRec 72.84%, NetRec 70.45%
• DumRec performs best among all, however the differences are not large.
• DumRec, most of the time, guess the correct letter or a "reasonable" wrong one.
• NetRec and FnetRec are very similar when compare letter by letter, though in general the latter slightly outperforms.

• DumRec - Probably because the features considered are too low-level.  Better recognition requires the use of higher-level features (e.g., roles).
• NetRec and FnetRec - Style may interfere with recognition.

Summary author's notes:

• none