In the experimentation with several designs of multi-layer perceptron-type neural networks in problems of function settings, the results obtained in terms of the performance of a neural network with more than one hidden layer do not improve the performance of the network or the precision in the estimates with respect to a single layer network. It is possible that I am not designing the network correctly and I do not know the proportion to be taken into account in the number of neurons to be assigned in each hidden layer.
Some Issues with Neural Network:
In theory, it has been established that many of the functions will converge in a higher level of abstraction. So it seems more layers better results. When you increase layers, it may result in better accuracy but it’s not a thumb rule. You have to just test it with a different number of layers. Always start with single layer then gradually increase if you don’t have performance improvement.
Here are some way to improve the Training Performance:
https://d4datascience.wordpress.com/2016/09/29/fbf/
You can also check this link-
http://www.turingfinance.com/misconceptions-about-neural-networks/
Dear Dagnier Antonio Curra Sosa,
In fact there is no general rule concerning the determination of the optimal number of hidden layers, however experience shows that their number will not exceed the number of entries. Commonly researchers use a minimum of hidden layer for simple systems and their number can evolve for complex systems.
For more details and information about this subject, i suggest you to see links and attached files in topics.
- How to improve performance of Neural Networks – d4datascience ...
https://d4datascience.wordpress.com/2016/09/29/fbf/
- Is accuracy proportional to the number of the hidden layers and ...
https://www.quora.com/Is-accuracy-proportional-to-the-number-of...
- Fit Data with a Shallow Neural Network - MATLAB & Simulink
https://www.mathworks.com/.../fit-data-with-a-neural-network.htm...
- Function fitting neural network - MATLAB fitnet - MathWorks
https://www.mathworks.com/help/nnet/ref/fitnet.html
- Learning and Soft Computing: Support Vector Machines, Neural ...
https://books.google.dz/books?isbn=0262112558
Best regards
Dear Dagnier Antonio Curra Sosa,
Sometimes, when we use more than one hidden layer, it dont improve the accuracy. Some studies that i suggest are to do some preprocessing data and to choose a proper activation function (this influence in the backward processing calculation).
introduction to preprocessing:
https://www.neuralstatistics.com/en/preprocessing/
variations of activation functions:
https://en.wikipedia.org/wiki/Activation_function
Thank you all. I will read and try each of your suggestions and comment on them later.
Not necessarily Mr Dagnier Antonio Curra Sosa ; you must changing the number of hidden layers and the number of neurons in the hidden layers until you find a best architecture, and the best architecture is the one who minimize statistical comparison like MSE, MAE, and so on.
and this number is related to the complexity of the things that you hope to solving with ANNs.
There is some task for example you can deal with just 5 neurons in one hidden layer. So you must changing this numbers until you find what you seeking for.
- Badges
- Science topic
- Similar topics
- Social Psychology
- Attitude