Inconsistent equation

Issue #225 resolved
Howard Yanxon created an issue

In https://bitbucket.org/andrewpeterson/amp/src/2865e75a199a386dca2c3dc77b2d63b854bf0e8d/amp/model/neuralnetwork.py?at=master&fileviewer=file-view-default#lines-884 line 844, the equation is inconsistent with the R. Rojas book (the first equation in page 168).

Is there a reason if you guys exclude (o_j(2) - t_j)?

Comments (2)

  1. cheng_zeng1

    Hi Howard, I think you mean line 884. The reason is that the ‘N+1’ layer is not the end (total energy) of the whole neural network, but an intermediate output of the atomic neural network. Line884 is in the block of calculate_ohat_D_delta , which is used to calculate the derivative of atomic energy w.r.t. parameter. Actually, (o_j(2) - t_j) is included in the evaluation of the derivative of final loss w.r.t. parameter (see here). I think (amp_energy - actual_energy) is the term you asked, and denergy_dparameters is the derivative of atomic energy.

    Hope this helps.

  2. Log in to comment