0

Why is my network showing me a high error?
I need to follow these rules - 

multi-layer feedforward (2 inputs, 1 output);

the first input has 262144 values (from 0 to 256) and the second 262144 (from 0 to 1024).

I'm using only one hidden layer. My error is something like this:

Epoch: 1; Error: 2816810148.1; 
Epoch: 2; Error: 2814260288.59; 
Epoch: 3; Error: 2813602739.7; 
Epoch: 4; Error: 2813385229.99; 
Epoch: 5; Error: 2813308095.39;
Arnab Nandy
  • 6,472
  • 5
  • 44
  • 50
hector
  • 1
  • 2

2 Answers2

0

You should normalize your input for first and second column. Then de-normalize your output. Also consider scaling your second column to be closer in values to first column. It will give your better error surface.

Yuriy Zaletskyy
  • 4,983
  • 5
  • 34
  • 54
0

It would appear that your network is working exactly as intended. On every epoch, the error has come down. The network initialises with a random "guess" and then moves from there - it has no foresight into the answer it must generate. With 250K+ inputs of values that go into the 1000s, it's not difficult to imagine that your error would be so high (you haven't stated what your error metric is). It's a lot of data, it's going to take a lot of epochs (1000s minimum) to get something useful. Between epoch 1 and 2 your error, whatever unit that may be, has come down by 2.5 million points. I'd say that was some improvement. Have you left it running for a few hours to see whether it might be able to solve this problem within a reasonable amount of time?

roganjosh
  • 12,594
  • 4
  • 29
  • 46
  • Yes, indeed it should run 2000 epochs, but at epochs 16-20, the error stops decreasing. The output has 262144 values (from 0 to 256), and my error goal is "0.0001". – hector Jun 11 '15 at 02:00