IMM_tensorflow
IMM_tensorflow copied to clipboard
About equation
Hello, I read your excellent paper, but due to my lack of mathematical foundation, I encountered some problems in formula reasoning. Can you deduce in detail how Equation 3 to Equation 8 is implemented as a whole, and if so, thank you very much for your help.
What is the connection between the derivation of the equations in Section 3 and the Weight Transfer, L2 Transfer, and Dropout Transfer?
Sorry for late response.
The mean-IMM is a method that approximates a Gaussian mixture as a single Gaussian through alpha blending, in the equation 3-5. And the mode-IMM uses Laplacian approximation to approximate Gaussian mixture. In the process, Fisher information is used as a mode. (c.f. equation 6-8)
Weight Transfer, L2 Transfer, and Dropout Transfer are different methods for multi-task transfer learning.
Thank you for your assistance and for answering my questions.