- Title: Empirical optimization (add10)
- Goal: Given a data set, an architecture, and a goal training error
arriving at the goal training error as fast as possible.
- Milestones:
- Please go through the
add10 training data and
description of data .
Using a feed-forward neural network with
10-5-1 (input-hidden-output) units at each layer
bias weight at each hidden and output unit
sigmoidal hidden and linear output activation functions
train on all data 20 times,
starting from random initial weights
till a mean square (1/N * sum_i (f(xi)-g(xi))^2)
training error of 2.3 is reached.
Report us back:
"the 20 inital and final training errors",
"the 20 final networks" (see print_w in support.h)
and "how much time each run took, their mean and standard deviation"
(use /usr/bin/time command, report the real time)
on cco's piccolo.
- Further Issues:
- Tentative Schedule:
- References:
- The optimization algorithms from
experiment 3.
- "Neural Networks for Pattern Recognition" by Chris Bishop.
- Caretaker: Zehra Cataltepe