Sorry for the naive question,
Why do we use neural network if we are going to apply logistic regression? Isn’t the logistic regression will perform same as neural network? I am failing to get the intuition behind using neural networks.
Using a neural network in helps in understanding more complexities within the data. If you are trying to make a prediction, the neural network will have a better understanding of how the features actually relate to the output. And it depends on the particular dataset. There are datasets where you are better off using logistic regression to accomplish the same task.
It’s just the same reason why we do not use simple linear regression for a classification task. Because, the linear model cannot understand the feature relationships that will lead to a categorical prediction.
Thank you for the explanation. How to identify which model to use? Also, when you say “more complexities within data”, what do you mean by that? Do you mean randomness, size of data etc. Maybe it comes by solving more problems using ML.
Let’s say we want to see whether images contain cats or other animals. This is a case of unstructured data (images) and this is where neutral networks thrive. This is what I mean by complexity. Contrast this to predicting whether a customer will default on a bank loan when you are given a table of details related to the customer e.g. income, debt, age etc. This second case is structured data (tabular data in rows and columns) and a simple logistic regression may suffice.
This is just a simple analogy. In general, while dealing with unstructured data (audio, images and the likes), neural networks are generally preferred.
However, you could still use neural networks for structured data too. It all depends on the data you have and your preferences.
Experimentation is the most common method, guided by your personal experience.
Add just enough complexity to get “good enough” results.
1 Like