next up previous contents
Next: Audio-Visual Speech Recognition Up: Integration Models of Previous: Conclusion

Taxonomy of the integration models

We will here make an attempt to classify the above presented architectures under a synthetic form and to rely them to the general models of cognitive psychology. To do so, we will organize the architectures through three questions:

  1. Does the interaction between modalities imply or not a common intermediate representation?

    The answer allows a first opposition between the DI model (for which it is NO) an the other architectures.

  2. If there is an intermediate representation, does it rely on the existence of prototypes or of early classification processes, i.e., at a symbolic level? In other words, is it a late or an early integration (see [348, page25,])?

    We will say that integration is late when it occurs after the decoding processes (SI case), even though these processes give continuous data (as with the FLMP). Otherwise, integration is early when it applies to continuous representations, common to the two modalities, and which are obtained through low-level mechanisms which do not rely on any decoding process: It is the case with the RD and MR models.

  3. Is there at last any dominant modality which can give a common intermediate representation in an early integration model (RD)? Or, is this common representation amodal (such as in the MR model)?

    Given this, it leads to the taxonomy presented in figure C.6 .

Figure C.6 : Taxonomy of the integration models

Esprit Project 8579/MIAMI (Schomaker et al., '95)
Thu May 18 16:00:17 MET DST 1995