Machine Learning Crash Course: Generalization

Key Takeaways

imagine you're a visitor from a far away Galaxy you're excited to go sightseeing and make new friends but you don't speak the language and your translator app isn't much help how are you going to communicate you wonder then you have an idea you find a bustling local community and listen for a few months so you can learn the language from native speakers after a few months of language immersion you finally feel ready to communicate uh-oh what went wrong clearly your language skills are not as good as you thought they were they didn't translate to the real world machine learning models can have the same problem when you train a model on data you're teaching it to perform well on that specific set of data but this training data is just a small subset of real world data if the model makes highquality predictions on the training data but much lower quality predictions on brand new data then the model didn't generalize in this case we say that the model has overfit the training data it has been tuned so closely to the specific patterns of the data it's already seen that it can't identify patterns in new data realizing your mistake M AK overfitting to the speech patterns of only a small group of people you begin your language studies again but this time you broaden your approach listening to a wide variety of speakers your communication skills now successfully generalized to brand new contexts and you land an exciting new job sharing your knowledge with others