Maximum PC - - R&D -

As you dig deeper into ma­chine learn­ing, you’ll find your­self tun­ing how the model is trained and re­trained to im­prove its per­for­mance. You do this by al­ter­ing things such as the num­ber of lay­ers in the net­work, and the num­ber of train­ing steps. These are called hyperparameters.

In ma­chine learn­ing, there are pa­ram­e­ters and hyperparameters. Pa­ram­e­ters are the data that the model turns into a prediction, such as pix­els in the im­ages, or words if we’re do­ing text clas­si­fi­ca­tion. Hyperparameters af­fect how well the model per­forms once trained.

For ex­am­ple, we’re set­ting “how_­many_­train­ing_steps” to 500, but it could be that 1,000 train­ing steps would cre­ate a more ac­cu­rate model, or maybe 300 is enough. Play about with the num­ber. Re­train the net­work us­ing fewer train­ing steps, then see whether the con­fi­dence in iden­ti­fy­ing the im­age per­forms as well. Then re­train with more steps, and com­pare your find­ings.

If you get re­ally into this, you can plot the per­for­mance of your model against the num­ber of trained steps to find the op­ti­mum num­ber. Do­ing this is known as “hy­per­pa­ram­e­ter op­ti­miza­tion.”

Newspapers in English

Newspapers from Australia

© PressReader. All rights reserved.