Search code examples
machine-learningneural-networkleast-squares

α-LMS‬ algorithm setback issue


I'm using α-LMS algorithm to make a classifier for MNIST dataset.

I only update the weights when predicted output does not match the desired output and here's how the weights are updated:

Weight update formula

I expect the residual error to go down (and accuracy to go up) on each iteration (shouldn't I?).

Although the overall accuracy seems to be going upward, on some iterations the algorithm sets back.

image

I need to know if it is normal for the setbacks to happen in this algorithm and if it is why so?


Solution

  • If the data could be represented by a formula, it would be feasible to find that formula. When an algorithm is needed (no formula), the speed of convergence and accuracy depends strongly on data. If you are able to use several datasets you'll see different graphs.