Hyperparameter: Difference between revisions

*[[Learning rate]]: This hyperparameter controls the step size used to update parameters in a model during training. A high learning rate may cause the model to converge quickly, but may also overshoot its optimal solution and produce suboptimal performance. Conversely, a low learning rate could cause slow convergence or lead to suboptimal solutions being found.

*Number of [[Hidden layer]]s: This hyperparameter determines the number of layers in a neural network model. A deep network with many hidden layers can capture complex features and patterns in data, but may also be susceptible to overfitting. On the other hand, a shallow network with few hidden layers may be easier to train but may not capture all pertinent information present in the dataset.

*[[Regularization]] strength: This hyperparameter determines the strength of a penalty term used to prevent overfitting in a model. A high regularization strength can help avoid this problem, but may also lead to underfitting the training data. On the other hand, low regularization strengths may provide good fit with training data but may not generalize well to new data sources.

* Grid Search: This method involves evaluating a model's performance against a grid of hyperparameter values. While this can be time-consuming for models with many parameters, it is simple to implement and usually produces promising outcomes.

* Random Search: This technique involves randomly selecting hyperparameters from a distribution and evaluating the model's performance for each set. While it may be faster than grid search in models with many hyperparameters, it does not guarantee finding the optimal set of hyperparameters.

* Bayesian Optimization: This technique involves creating a probabilistic model of the hyperparameters and using it to select the most promising set for evaluation. It can be more efficient than grid search or random search, often producing good results with fewer evaluations.