Neural Turing Maching:
-
Memory : M N × D M^{N\times D} MN×D
-
Controller
-
Write Head
w t ( i ) = = e x p ( β K ( M t ( i ) , q ) ) ∑ k e x p ( β K ( M t ( i ) , q ) ) w_t(i) = = \frac{exp(\beta K(M_t(i), q))}{\sum_{k}exp(\beta K(M_t(i),q))} wt(i)==∑kexp(βK(Mt(i),q))exp(βK(Mt(i),q))
K ( u , v ) = u ⋅ v ∣ ∣ u ∣ ∣ ⋅ ∣ ∣ v ∣ ∣ ∣ K(u,v) = \frac{u\cdot v}{||u||\cdot||v|||} K(u,v)=∣∣u∣∣⋅∣∣v∣∣∣u⋅v
M t + 1 ( i ) = M t ( i ) ⋅ [ 1 − w t ( i ) ⋅ e i ] + w t ( i ) ⋅ a i M_{t+1}(i) = M_t(i)\cdot[1 - w_t(i) \cdot e_i] + w_t(i)\cdot a_i Mt+1(i)=Mt(i)⋅[1−wt(i)⋅ei]+wt(i)⋅ai
- Read Head
w t ( i ) = = e x p ( β K ( M t ( i ) , q ) ) ∑ k e x p ( β K ( M t ( i ) , q ) ) w_t(i) = = \frac{exp(\beta K(M_t(i), q))}{\sum_{k}exp(\beta K(M_t(i),q))} wt(i)==∑kexp(βK(Mt(i),q))exp(βK(Mt(i),q))
K ( u , v ) = u ⋅ v ∣ ∣ u ∣ ∣ ⋅ ∣ ∣ v ∣ ∣ ∣ K(u,v) = \frac{u\cdot v}{||u||\cdot||v|||} K(u,v)=∣∣u∣∣⋅∣∣v∣∣∣u⋅v
r t = ∑ i = 1 N w t ( i ) ⋅ M t ( i ) r_t=\sum_{i=1}^N w_t(i)\cdot M_t(i) rt=i=1∑Nwt(i)⋅Mt(i)