What is the best way to change the neural network?

Question

Neural networks

Good afternoon. My friend and I are doing a small pet-project.
The task was formulated as follows.
There are 7 object classes. Each has 437 binary features. There is a training sample for ~400 cases. Due to the nature of the domain, the number of instances of each class is different. It is required to make a classifier that, based on the newly entered data, will determine to which class the new object belongs.
Our actions
We decided to make ANN. The sample was trimmed artificially (where there were very few cases, they were dug up).
INS we have with one layer. And everything seemed to work out fine, if not for one thing:
Problem
One of the classes is defined constantly incorrectly. Those. on six of them, the percentage of hits is 94-96%. We are satisfied. But one by one, the INS stubbornly lies. And for this class, initially the number of cases was heaps.
We rechecked the sample - no, according to the signs, everything is correct.
Retrained the network - the result is the same.
Probably it is necessary to change architecture of a network. And here we have a question. On what? Can you advise something?

Answer 1

> 437 binary features. There is a training sample for ~400 cases.
This means that your neural network will be retrained. For any machine learning, the number of training rows must be much greater than the number of columns. If you reduce the number of columns, this (in my opinion) will not change anything drastically in this case.
In a neural network, at least one hidden layer is better, if possible, increase the number of hidden layers. Try different layer configurations ("barrel", "glass", "rectangle"). You can try a discrete number (class number) in the output layer, or you can try in the output layer the number of outputs = the number of classes (all zeros, except for one unit opposite the desired class).