Energy-based models

• Use a neural network to assign an “energy” to each possible data- vector. Then define probabilities in terms of energies. – Learning is done by manipulating energies.

It is easy to lower the energy of d

We need to find the serious rivals to d and raise their energy. This seems hard.