Both Deep Learning and Symbolic Regression can be used to solve machine learning problems, but the approaches are entirely different. Here we will explain which problems are most suitable for each method.
What is Deep Learning?
This is the type of machine learning which is mostly used in self-driving cars, artificial intelligence, speech recognition, etc.
One of the biggest advantages of Deep Learning is its ability to automatically learn the relevant features present in the data, resulting in accurate models even in challenging problems with large numbers of dimensions.
Deep Learning can achieve this kind of sophisticated representation of the data by employing many different layers of hidden neurons.
Let’s take a look at an example: say you have images of cats and dogs and want to automatically classify those images. Deep learning is capable of turning the RGB pixel data into meaningful features to be used in the classification. The problem would be hopeless without this automatic feature extraction since the data from a few individual pixels is not enough to classify the images.
What is Symbolic Regression?
Symbolic regression can be thought of as the analog of regression analysis, but with the functional form of the model being open. The idea of the method is to try to discover by trial and error explicit mathematical formulas that connect the variables.
Typically nonlinear base functions like exp(x) or cos(x) are used to try to construct the formulas, resulting in models capable of capturing very subtle correlations between the variables.
When to use Deep Learning vs Symbolic Regression?
As a rule of thumb, deep learning is most useful for problems in spaces with large number of dimensions, like image classification — a simple 500px image is defined by 750,000 numbers, and formulas that combine this many numbers into meaningful results would have to be large and cumbersome.
Symbolic Regression on the other side is very useful for problems in fewer dimensions. The method will automatically pick the most relevant input variables to try to predict the target variable, resulting in simple and robust models that may reduce the chances of overfitting.
Conclusion
Both Deep Learning and Symbolic Regression are powerful machine learning methods. Deep Learning is suitable for problems that are hard to directly tackle using naive algorithms due to the large number of dimensions involved. A good example is the problem of identifying objects in images. Symbolic regression on the other hand is a very powerful method for discovering simple and explicit relationships between variables.