Hello! The first post on Habré and immediately a hardcore topic on the topic of the day. I think that many developers of artificial intelligence for solving applied problems wondered which neuron architectures are most effective in the context of specific tasks. I'll make a reservation right away that the examples given were developed by the staff of the University of Artificial Intelligence. But I, as a participant in their intensive, was lucky to test their architectures and collect useful statistics on their effectiveness.
1. Recognition of handwritten numbers
Let's start with the simplest architecture. This mesh consists of one input and 3 fully connected layers:
This simple grid showed very good results during training. The accuracy on the training sample was 99.4%, while on the test sample it was 98.5%. And this in 2.57 seconds! Go ahead.
2. Recognition of the car brand
The second grid is heavier, but the task was more ambitious. For the experiment, photographs of three brands were taken - Renault, Mercedes and Ferrari. The model consists of the following layers:
"" , . 2D, 2D, , . 76,7%, - 73,6%. - 1,7 .
3.
- . :
. 100,0%, - 99,9% (!). . - 0,7 . .
4.
, - , . - , . .
. , , , 1D, 1D, . 82,7%, - 85,1%. , . . - 0,16 .
5. ...
, ? . , , . :
, . - , PSP . - , . , 2 . - (3 ) (2 ). 99,8%, - 99,8%. , . 4,7 .
The given examples of architectures during the tests showed good results and can be applied in solving practical problems. For each of the models, about 20-30 tests were carried out to change their parameters. Perhaps in the following publications I will give detailed testing ranges for the presented models. Thank you for attention!