14. Learning: Sparse Spaces, Phonology

Visual recognition to sound production

Words can be classified in tables representing the 14 features of each sound, notably if they are “voiced” or not. With such a matrix of sounds, a “speaking machine” can be built to produce words and sentences.

Example: a machine pronouncing plurals

The machine processes from recognition through different characteristics of words (verbs, nouns, plurals…) and constraints (pronunciation of plural) to produce sounds. The machine’s input comes from visual recognition and its output is sounds to mimic a voice to produce the right sound, “z“, “s” or “e-z“.

  1. The vision apparatus recognizes objects
  2. The information flows to the meaning register, labeling the objects, as well as through characteristics register to define if they are plural or not
  3. The information from the meaning register also flows through the characteristics register to define if they are nouns or verbs, and to a buffer which stores the beginning of the word letters while waiting to know if a “z” or “s” or “e-z” plural sound needs to be added.
  4. The buffer letters moves to be processed by the pronunciation apparatus, waiting to be processed by the plural constraint.
  5. The word is send to the plural constraint processor.
  6. If the word is plural and the last letter is voiced, the constraint adds a “z” sound, if not voiced an “s” sound is added. If the word is not plural, nothing is added, ending the word’s pronunciation.

Such machine is called a propagator as connections are reversible: by hearing a plural word, the machine can figure where are many of the objects to be seen by the visual apparatus.

Building phonological algorithm

  1. Collect + and – examples
  2. Pick and seed positive
  3. Generalize by churning useless information where + and – examples are not differentiated
  4. Quit when cover a negative example, or go back to step 3 and generalize more

Sparse space

The plural machine works because in a high dimensional sparse space, a hyperplane can easily separate two sets of examples, as phonemes constitutes a 14-dimension space.

Generalizing to problem solving with AI

  1. Define the problem
  2. Devise a representation suiting the problem
  3. Determine an approach / method
  4. Pick a mechanism or devise an algorithm
  5. Experiment

Troubles ensues when people focus on a specific mechanism rather the problem at hand.

The right representation does the following:

  • it makes the problem explicit
  • it exposes constraint
  • it generates localness to solve the problem: the answer can be pinpointed rather than gathered from various sources, areas, places, etc.