- For regular data with output --> Use Standard Neural Network
- For image data, data with spatial relations --> Use Convolutional Neural Network
- For sequential data, for signal data like Audio data or data where current value depends on previous --> Use Recurrent Neural Network
- As per requirements, use hybrid architectures, like CNN + NN, RNN + NN, CNN+RNN
- Neurons are poor processors, slow and often unreliable. This disadvantages can be overcame by using large number of neurons in parallel, with many connections. These can adapt changes.
- In conventional computers, operations are rule based performance sequentially. Also they perform detailed specification. Hence changes in environment may lead to trouble. So such computers are appropriate for performing well defined rule based information processing tasks in stable and safe environment. Neural information processing unit with real world tasks where data is often messy, uncertain and has virtually no perfect solution exist.