In machine learning sometimes we need to know the relationship between the data, we need to know if some predictors or features are correlated to the output value, on the other hand sometimes we don’t care about this type of dependencies and we only want to predict a correct value, here we talking about inference vs prediction.

Prediction

The most intricated algorithms in machine learning for example those associated with deep learning are very advanced algorithms that usually ingest a lof of data to create inner patterns, for example a functional neural network will need a lot of data if it wants to be highly reliable, when you train a neural network with backpropagation a lot of things are happening inside the structure of the neural network but you don’t know exactly what is happing, what we know is that the weights are being balanced and corrected with each new input, the final neural network configuration can be associated with a mathematical function that can describe the base original input data with the associated outputs, this function is a black box because we don’t know the implementation, but in some cases we don’t care about the implementation because what we want is that the function outputs a very good result. In this case, we are saying the neural network is trained and can predict most of the values we provide it. It is called prediction because we don’t have knowledge what kind of rules produce the output value of the neural network, for us this is a black box that outputs a good value most of the times, most of the deep learning algorithms out there output predictions because of their complexity but at the end we are fine with that.

Inference

On the other side it is good to understand why a machine learning algorithm is giving us a specific result, this is to be able to understand the relationship between the multiple variables that our input data have in common, this is great because we can put algorithms to work for us to understand the data, and even better we can have insights of the relationship between one variable or another. The reality is that most of the complex and most powerful algorithms will not give us inference about the data because they handle a lot of input parameters that becomes very complex to understand using statistical tools, this is unfortunate because it will be great to have a generated rule about an autonomous car driving, on the other side is util for another kind of problems with fewer parameters.