The purpose of machine learning is to find a function through data that can map inputs to outputs. For example, given an image, this function can map the input image to corresponding descriptive text.

How do we find this function? The process of machine learning first involves setting the form of the model and the parameters to be used. Generally, models mostly use neural networks, and the parameters within are referred to as weights. A large model may have many parameters, some reaching billions. To find good parameters is beyond human capability, so the task is left to machines to compute. This computing process is known as training the model.

I tested the Gemma model this morning. My computer has 32G RAM and an RTX 2070 graphics card (8G). I downloaded the 2B model. There are un-tuned and instruction-tuned versions, each of which is about 3.9G after unpacking. I executed it in CPU mode, but it crashed after a while. I guessed it may be because there is not enough memory. Then I used GPU mode to execute, the results were as follows: