Iteration: Difference between revisions

Latest revision as of 21:23, 17 March 2023

See also: Machine learning terms

Introduction

In machine learning, Iteration is when a model updates it's parameters (weights and biases) one time during training. The number of [example]]s the model processes in each iteration is determined by the hyperparameter batch size. If the batch size is 50, the model processes 50 examples before updating it's parameters - that is one iteration.

Machine learning involves iteration, which is the process of optimizing parameters in a model to enable accurate predictions on new data. This involves adjusting the parameters based on errors made during training on a training dataset. By repeating this process multiple times, the model learns from its errors and improves its accuracy.

One common application of iteration in machine learning is gradient descent, an optimization algorithm designed to find the minimum cost function. In gradient descent, the model's parameters are updated iteratively based on the gradient of the cost function with respect to the parameters.

What Happens in an Iteration?

In training a neural network, a single iteration includes:

A forward pass to calculate the loss on a single batch of data.
A backward pass (backpropagation) to modify the network's parameters based on the loss and the rate at which it is learning (learning rate).

Types of Iterations

Machine learning often employs several types of iterations, such as:

Stochastic gradient descent (SGD): when each iteration uses only 1 example of the training data. After processing just 1 example, the model updates its weights and biases. While it is fast, SGD can be unstable.
Mini-batch gradient descent: when each iteration uses a randomly chosen subset of training data to balance speed of convergence with stability in the optimization process.
Batch gradient descent: when each iteration uses all of the training data. This form of gradient descent offers stability but may be computationally expensive for large datasets.

Explain Like I'm 5 (ELI5)

Iteration in machine learning is like making educated guesses to find the correct answer. Imagine playing a guessing game with friends and they tell you if your guess is too high or low; using that information, you can use it to make an even better guess the next time around. This process of making one guess and using feedback for further refinement is known as iteration.

Machine learning relies on iteration to help computers learn from data. The computer begins with an initial guess about how to make predictions, and then updates its guess according to how well it did. This cycle continues until it gets as close to the right answer as possible.

@@ Line 7: / Line 7: @@
 One common application of iteration in machine learning is [[gradient descent]], an [[optimization algorithm]] designed to find the minimum [[cost function]]. In gradient descent, the model's parameters are updated iteratively based on the [[gradient]] of the [[cost function]] with respect to the parameters.
-==Types of Iterations in Machine Learning==
+==What Happens in an Iteration?==
-Machine learning often employs several types of iterations, such as:
+In training a [[neural network]], a single iteration includes:
-Stochastic Gradient Descent (SGD): SGD allows the model to be updated using a randomly chosen subset of training data instead of using the entire dataset. This enables it to make rapid progress towards minimizing its cost function; however, this may introduce noise into the optimization process.
+#A forward pass to calculate the [[loss]] on a single [[batch]] of data.
+#A backward pass ([[backpropagation]]) to modify the network's [[parameters]] based on the loss and the rate at which it is learning ([[learning rate]]).
-Mini-batch Gradient Descent: Mini-batch gradient descent involves updating a model using a randomly chosen subset of training data to balance speed of convergence with stability in the optimization process. This minimizes errors associated with model updating.
+==Types of Iterations==
+Machine learning often employs several types of iterations, such as:
-Batch Gradient Descent: With batch gradient descent, the model is updated using all of the training data. This form of gradient descent offers stability but may be computationally expensive for large datasets.
-==Explain Like I'm 5 (ELI5)==Iteration is the process of doing something over and over again, such as playing a game of tag to improve at it. In machine learning, iteration refers to when a computer program keeps trying to improve its accuracy in predicting things by altering its settings each time something goes wrong (like getting tagged in tag), then trying again. Over time, this helps the program get better at making predictions with increased practice.
+#[[Stochastic gradient descent]] (SGD): when each iteration uses only 1 [[example]] of the [[training data]]. After processing just 1 example, the model updates its weights and biases. While it is fast, SGD can be [[unstable]].
+#[[Mini-batch gradient descent]]: when each iteration uses a randomly chosen subset of training data to balance speed of [[convergence]] with [[stability]] in the optimization process.
+#[[Batch gradient descent]]: when each iteration uses all of the training data. This form of gradient descent offers [[stability]] but may be computationally expensive for large datasets.
 ==Explain Like I'm 5 (ELI5)==
@@ Line 22: / Line 24: @@
 Machine learning relies on iteration to help computers learn from data. The computer begins with an initial guess about how to make predictions, and then updates its guess according to how well it did. This cycle continues until it gets as close to the right answer as possible.
-==Explain Like I'm 5 (ELI5)==
-Imagine you have a large bag of candy and you want to find the yummiest treat within it. To do this, you could taste each candy one by one and decide if it's tasty or not - this process is known as "iteration".
-Machine learning uses a similar process to find the optimal solution to problems. Instead of candy, we provide instructions that the computer can follow to make a decision or solve an issue. And just like tasting each candy to find which is yummiest, this computer runs through these instructions many times with small modifications each time until it finds the ideal solution - this iterative process being known as "iteration".
-[[Category:Terms]] [[Category:Machine learning terms]]
+[[Category:Terms]] [[Category:Machine learning terms]] [[Category:not updated]]