Oleg Zabluda's blog
Monday, April 17, 2017
 
Learning To Learn Using Gradient Descent (2001) Sepp Hochreiteri, et al
Learning To Learn Using Gradient Descent (2001) Sepp Hochreiteri, et al
"""
The training data for the meta-learning system is a set of sequences {s_k}, where sequence s_k is obtained from a target function f_k. At each time step j during processing the kth sequence, the metalearning system needs the function result y_k (j) = f_k (xk (j )) as a target. The input to the meta-learning system consists of the current function argument vector x_k(j) and a supplemental input which is the previous function result y_k (j -1). The subordinate learning algorithm needs the previous function result y_k (j -1) so that it can learn the presented mapping, e.g. to compute the subordinate model error for input x_k (j -1). We cannot provide the current target y_k (j) as an input to the recurrent network since we cannot prevent the model from cheating by hard-wiring the current target to its output.
"""
http://snowedin.net/tmp/Hochreiter2001.pdf
http://snowedin.net/tmp/Hochreiter2001.pdf

Labels:


| |

Home

Powered by Blogger