I see a correspodence between the GEN and EVAL components in OT and the Generative Model and the Discriminatory Model in GANs. I wonder if this goes beyond a superficial similarity. Do you believe there are further insights into GANs we can squeeze out of OT ?
.
structurally there are some similarities, in the sense that there is a "Gen" module and an "Eval" module evaluating the productions of "Gen" ... but the same could be said of the "actor-critic" formalism in reinforcement learning and the correspondence seems superficial to me
a first main difference regarding "Gen" :
- in OT, "Gen" takes a motif and generalizes (makes some modification to) it ; ergo, "Gen" has direct access to the initial database
- in GAN, the "Gen" module never accesses the original database : its only "contact" to this database is through the feedback of the "Eval" module
another difference is in the feedback signal from "Eval" :
- in OT, there is no real feedback to Gen from Eval : among the motifs produced by Gen, Eval finds the best motif by striking a balance between "markedness" (is the generalized motif well-formed, utterable ... ?) and "faithfulness" (is the generalized motif close enough from the original motif ?)
- in GAN, the feedback is indeed sent to Gen and measures the abilty of Eval to "see" a difference between the motifs produced by Gen and the original motifs (that only Eval has access to) ; moreover, this feedback is the only training information Gen gets
.
now, being a specialist in neither field, i'll be glad to be corrected !
.
Goodfellow et al. (2014) do reference a paper by Smolensky. It's of his earlier work (1986) on what ended up as Harmonic Grammar (1991), though. Smolensky later went on to co-formulate Optimality Theory (1993). So, Goodfellow et al. had picked up on a strand of reasearch that eventually gave birth to OT. Is it that they simply didn't follow it through, missing out on insights from the newer paradigm? Or is it that Smolensky moved to a directon that no longer complied to their vision of the GAN?
.
this reference seems just there to acknowledge the introduction of RBMs with nothing related to "adversarial training" ; the following paragraph just sums up the difficulties associated with training such models
.
- Badges
- Science topic