next up previous contents index
Next: Generalization: Up: Stability and generalization Previous: Stability and generalization   Contents   Index

Stability:

Some authors use the term stability to refer to the following property: that a DTRNN is exhibiting stable behavior (in the sense of being a neural language recognizer or a language transducer) when outputs are within the regions of output space assigned to the corresponding symbols for strings of any length, as discussed in section 4.2.

In general, trained DTRNN are ``stable'' only for strings up to a given length; this is due to the nature of the internal representation assumed. For example, if the DTRNN is trained to behave as a finite-state machine instability means that the regions of state space visited by the DTRNN corresponding to the FSM to be inferred are not disjoint and merge as they grow with string length; as a consequence, they do not clearly map into the regions $Y_m$ of output space assigned to the desired outputs.

As shown in chapter 4, DTRNN may be constructed so that they behave stably as FSM.


next up previous contents index
Next: Generalization: Up: Stability and generalization Previous: Stability and generalization   Contents   Index
Debian User 2002-01-21