Imagine you’re trying to train a dog to perform a trick, like catching a frisbee. You don’t know the best method to train the dog, so you start with one method and see how it works.
- Start with a training method.
- See if the dog is getting better or worse at catching the frisbee.
- Adjust your training method a little based on the results. If the dog is getting better, keep doing more of what you’re doing. If the dog is getting worse, try something a little different.
- Repeat steps 2 and 3 until the dog is really good at catching the frisbee.
Stochastic Gradient Descent (SGD) does the same thing, but instead of training a dog to catch a frisbee, it’s used to “train” a computer model to perform a task better.
The “training methods” are all the possible ways the model could work, the “frisbee catch” is a measure of how well the model is performing its task, and the “adjustments” are small changes to the model.
SGD’s goal is to find the best model by starting with a random one, checking how well it performs, making a small adjustment to improve it, and repeating this process many times. So, just like you’d adjust your methods to train a dog, SGD adjusts the model to make it perform better.
« Back to Glossary Index