The model may not converge for various reasons

[rifat28dddd]

Can a model "not converge"
There are situations when convergence cannot be achieved for some reason. This is called divergence. With divergence, the values ​​of the objective function do not stop changing - the losses do not reach a minimum.

If a model diverges, no matter how much it is trained, it will never improve its accuracy to a sufficient level. Divergence usually indicates that the model itself, the method, or the training parameters need to be changed.


because of the characteristics of the data it is trained on - say, it is not scaled or normalized, it contains outliers or noise;
due to incorrectly chosen loss function;
due to inappropriate hyperparameters of the model, such malta telegram data as the training step - if it is too large, the model will diverge.
Often, gradient optimization methods, such as gradient descent, are used when training models. In this case, the cause of divergence may be too large a step of the gradient - a vector that shows how quickly the values ​​of the function change. And if the gradient of the loss function is zero, the model will never converge - this means that the value does not change.

The most difficult situation is when the algorithm does not converge due to architectural errors. This sometimes happens with deep neural networks. For example, divergence can occur due to the lack of batch normalization or an inappropriate activation function.

Due to architectural errors, function gradients can decay or explode during training. When decaying, gradients approach zero, and training slows down significantly. When exploding, the gradient grows sharply, and further training becomes impossible.

Divergence is dealt with in different ways depending on the cause. To avoid it, an ML specialist should:

carefully prepare data before training;
correctly select loss functions depending on the task at hand;
use various optimization methods.