Are Neural Networks hard to build for custom data

Data Science, AI and ML
#1

are Neural Networks hard to build for custom data

#2

Yes because of its internal network

#3

We can reuse most of the already implemented networks based on our usecase. Data preprocessing and model hyper parameters and weights are the important thing to change

#4

The structure of neural network is discrete. You can’t have a part of a neuron. Neural networks aren’t very good at optimizing discrete things. The closest they come to this is classification. You could try to decide the structure of the neural network for a problem using another neural network for classification and it’s possible to have a little bit of success maybe, but you run into a lot of problems.

  1. The input. It’s hard to find a reasonable input that will make sense in a sufficiently wide class of problems that you can train the network.
  2. Complexity. You’re asking a neural network to compare the performance of other neural networks. To do that, its hard to get around having it simulate them, in which case actually running the simulations would work better.