The Machine Learning model building involves intensive and repetitive computations involving tensors. These require intensive computing resources. As a regular compiler would provide the optimizations at the local level, it does not generally produce a fast execution code.
Theano first builds a computational graph for the entire computation. As the whole picture of computation is available as a single image during compilation, several optimization techniques can be applied during pre-compilation and that’s what exactly Theano does. It restructures the computational graph, partly converts it into C, moves shared variables to GPU, and so on to generate a very fast executable code. The compiled code is then executed by a Theano function which just acts as a hook for injecting the compiled code into the runtime. Theano has proved its credentials and is widely accepted in both academics and industry.