In the Demand Prediction topic, it was mentioned that each neuron is doing a logistic regression computation. That output is passed onto the neuron in the next layer.
Can some neurons do linear regression and others logistic regression? Are there such use cases?
How the neurons act is controlled by the cost function.
If you’re using a logistic cost function, then all of the neurons in the output layer will have their weights updated as appropriate for logistic regression.
You can specify different activations for each hidden layer. This is actually very common. But the whole layer would typically have the same activation.
You could have a compound deep learning network where you could split a layer into two portions, and have different activation functions for each part. But you’d need a very good reason for using that type of complexity.