Signal level artificial neural network
by (c) Ville Turjanmaa, Univ.Pro-Seminar
Transition recognition neural network is a pattern learning and recognizing network, which can be applied to notated music and other patterns. The network shows hierarchial dependencies between occurred patterns. Transition ( TR ) neural network collects data without loss of information.
Can be applied to any time-related pattern.
Uses unsupervised learning.
Shows time-related hierarchial dependencies.
No loss of information.
Example caseThis document forms a TR -network from the following example. Chords can be used instead of single notes.
Information from the note sequence1) Quantize level
Quantize level is the period of shortest event in score. In the above example, quantize level is 1/16 note.
2) Number of input cells for the network.
The above example ( GEDE-CEDE ) uses four different notes, generating input cells G , E , D and C.
Transition connections and recognizing cellTransition example for first two notes in score.
A transition cell has two input nodes ( A and B ), output line and internal state ( 0 or 1 ). When signal arrives to node A, the internal state raises from 0 to 1. When input node B receives a signal while the internal state is 1, the cell fires and returns to internal state 0. When receiving a pair ( ie. 'CC' ), the first pair raises the internal state to 1, second pair fires the cell and keeps the state at 1, since other pairs could follow. All other cases should zero the cells internal state.
When firing, all cells in TR-network generate one impulse to the next level. The input data should be synchronized with hidden level cells. When forming a network, only one cell is generated for each transition.
Impulse response table for transition recognizing cell:
A B State -> State Out 0 0 0 0 0 0 0 1 0 0 0 1 0 0 0 0 1 1 0 1 1 0 0 1 0 1 0 1 1 0 1 1 0 1 0 1 1 1 1 1
Transition neural network
When any of the first hidden level cells fire, the input is a two note sequence from the example. When any of the second hidden level cells fire, the input is a three note sequence from the example.
Transition mapAll occurring patterns ( 25 ) with hierarchial dependencies
ConclusionTransition neural network can be used as a general tool for notation. TR-network learns and recognizes patterns without loss of information. The structure of formed network shows all dependencies between patterns. Both learning and recognizing takes place in same cell.
In addition to notation, transition neural network ( TRNN ) can be applied to other time-related data streams, like sentences and functions.
DownloadsMs-dos program which forms TR network: TR.ZIP
Notes for Bach BWV 847.1 Page 1 and Page 2