Given a business problem, there is no hard and fast rule to determine the exact number of neurons and hidden layers required to build a neural network architecture. The optimal size of the hidden layer in a neural network lies between the size of the output layers and the size of the input. However, here are some common approaches that have the advantage of making a great start to building a neural network architecture –
To address any specific real-world predictive modeling problem, the best way is to start with rough systematic experimentation and find out what would work best for any given dataset based on prior experience working with neural networks on similar real-world problems. Based on the understanding of any given problem domain and one’s experience working with neural networks, one can choose the network configuration. The number of layers and neurons used on similar problems is always a great way to start testing the configuration of a neural network.
It is always advisable, to begin with, simple neural network architecture and then go on to enhance the complexity of the neural network.
Try working with varying depths of networks and configure deep neural networks only for challenging predictive modeling problems where depth can be beneficial.
It depends on the associations you want it to represent. One way to understand what’s going on is the ``tiling algorithm" of Mézard and Nadal (the paper’s on ResearchGate, but it suffices to input tiling algorithm Mezard Nadal in one’s favorite search engine).
I completely agree that there is no "one size fits all" Theoretical Solution yet for this problem. That being said, I would also say that since there is no generally accepted Theoretical Solution, you are reduced to finding the "Best Fit" empirically validated answer. As in all problem solution finding, the first step in understanding and validating the solution answer selected is to define what is the criteria that defines your particular "Best Fit". One consideration is compute power available and for how long. This will define the maximum depth and width of the solution neural network layers. Note these criteria are not mutually exclusive. By that I mean you can sometimes trade off depth and width to get an acceptable solution. The problem then becomes a classic tradeoff analysis with many well known solutions. Now the problem becomes one of validation of the generated proposed solutions, and that leads to analysis of what validation tools you have available to "validate". [If you have known "use cases" with acceptable solutions that makes the problem of selection of the "Best Fit" a lot easier, but that all depends on your specific environment.]
- Badges
- Science method
- Similar topics
- Methodology
- Crystallography
- X-ray Diffraction