Hello everyone.
I am dealing with construction of extrapolation model with deep learning.
When the training dataset is a shape of (1,0,0,0), (0,1,0,0), (0,0,1,0), (0,0,0,1)
I want to build up a model that is able to reconstruct data that is shape of (1,1,0,0 ), (1,1,1,0).
Since the training data and test (or actually-using) data is different, it does not make sense for conventional facts related to deep learning.
The task now I am doing is, using the test data as a validation set of learning and finding the learning that minimize validation loss.
However, now I realize that the task costs too much resource and time. It can be same as doing some lottery.
Therefore, there should be some improvement for the methods.
If someone can give me some advice for this part, it would be a really big help.
Thank you so much.
Dear Mr. Park,
Put a 1234 weight on the digit counters and try to maximise the increment in the combinatorial value. Also try the weights -1 1 -1 1 and try the same. One of the weights should work with test data, at least there should be improvements.
Earl Prof. Dr. S.K.Mallick KBE QC ForMemEPS ForMemReS MemES MemISI MemMathFin (CA)
for RHMHM School of Mathematical Sciences Thinking
USA, Japan, India
Park,
Deep learning/NN is normally not suitable for extrapolation tasks. The network will be trained on the things that it has seen before, and recognizes those things only. It doesnt build any analytical model for the signals.
- Badges
- Science topic