Is a last layer of neurons in Neural Network a linear classifier? Can i consider a last layer of neurons a linear classifier regarding of inputs to the last neurons?
 A: Yes. If your last layer's activation is 'linear' or if there is no activation, then it is a linear regression. If the activation of the last layer is 'softmax', it is a logistic classifier.
Input to the last layer is basically features extracted by your neural network.
A: It's best to always ask: what is the NN being used for: function approximation (i.e. curve fitting) or class prediction?  For function approximation, you don't always need an activation function on the output side before the output nodes, and can just multiply hidden node outputs by connection weights on the output side (this would be like linear regression).  My experience is that when using a NN for function approximation, a lot of times a linear link (simple product of weights and hidden node outputs) works best.  But recall, if there are sparse regions of data over the range of the input features, an NN will break down easily since the data grid is not very uniform.  Thus, I commonly always implement LHS (Latin hypercube sampling) which simulates continuous input features uniformly over their range so the NN doesn't have convergence issues.  If you straightforwardly use a NN for function approximation and it bombs out, it's probably because you didn't employ LHS -- many papers on this.    
Regarding class prediction, for a 2-class problem, you only need one probability as an output node, and can use a logistic activation function on the output side.  But for k-classes, you would use softmax (Boltzmann distribution) which will give a probability for each class -- thus, if there a 4 classes, there will be 4 output nodes.  
