Automated search for optimal net architecture
Posted: Mon 6. Dec 2010, 14:59
Hello!
Just an idea how to help reduce hand-work time needed to find an optimal neural network architecture.
A user specifies the limits on the number of neurons in the hidden layer, e.g., 2-20. Then, the user specifies the limit on the number of hidden layers, e.g., 1-3. He also selects from the list of functions on the neurons to use which functions will be applied. Also the limit on the total number of trials that the program will run.
The program then generates various nets and trains them till some minimum error on the test (or train) sample is achieved. The best - trained - architecture is stored in the end of the procedure.
This idea is derived from my experience in using other NN packages. Seems helpful and time saving.
BR
Alexey
Just an idea how to help reduce hand-work time needed to find an optimal neural network architecture.
A user specifies the limits on the number of neurons in the hidden layer, e.g., 2-20. Then, the user specifies the limit on the number of hidden layers, e.g., 1-3. He also selects from the list of functions on the neurons to use which functions will be applied. Also the limit on the total number of trials that the program will run.
The program then generates various nets and trains them till some minimum error on the test (or train) sample is achieved. The best - trained - architecture is stored in the end of the procedure.
This idea is derived from my experience in using other NN packages. Seems helpful and time saving.
BR
Alexey