What is a disadvantage of an Artificial Neural Network?

Answer 1

1. They are black box - that is the knowledge of its internal working is never known

2. To fully implement a standard neural network architecture would require lots of computational resources - for example you might need like 100,000 Processors connected in parallel to fully implement a neural network that would "somewhat" mimic the neural network of a cat's brain - or I may say its a greater computational burden

3. Remember the No Free Lunch Theorem - a method good for solving 1 problem might not be as good for solving some other problem - Neural Networks though they behave and mimic the human brain they are still limited to specific problems when applied

4. Since applying neural network for human-related problems requires Time to be taken into consideration but its been noted that doing so is hard in neural networks

5. The Vapnik-Chervonenkis dimension or VC Dimension of a neural network which is a combinatorial parameter that measures the expressive power of a neural network is still not well understood

6. They are just approximations of a desired solution and errors in them is inevitable

7. Lastly I will add that they require a large amount training set to be trained properly and to give output(s) that would be close enough to the desired output but knowing what amount of training set is enough for a desired output would be totally dependent on the trainer itself - but yes its important that a very large training set is provided so that the neural network would have sufficient understanding of the underlying structure.

Answer 2

local minima

generalization/overfitting

hard to interpret