A couple more jargon terms on AI

If you have read the earlier article you probably have a flow going and you landed up over here. Here we will try and cover some common things that is a good to know today.

Training vs Inference

In the earlier blog we mentioned training a model which is nothing but passing data to a model to make better neural connections. Inference is when you use the trained model to do something. i.e. you use chatgpt to write something for you. Training is compute heavy vs inference which is cheap.

Fine tuning

The models like GPT are trained on generic data. In this sense, it is like the Internet. It has a lot of data. When you want to train the model to be used for a specific use case, we call it fine-tuning. Say you take the GPT neural network and tune it to respond to medical texts better. Basically, taking a generic model and making it specialise in a particular domain. The advantage is that you are not starting from scratch. You are just building on a pre-trained model. You saved some dollars.

How models learn

Now, even though we said we pass data and the model learns, there is a bit more to it. The data is actually passed in different ways. The ways are as such —

1. Supervised learning — in this type of learning, we pass the input and output of a sentence. For e.g. we train NN by saying “the cat sat on the _” and GPT has to guess and check if the answer is “mat” which we have also passed. This correct answer allows the NN to adjust its weights. A use case is email spam detection. We feed a lot of emails to the NN and tell it what is spam and what is not. And then it starts guessing if it is spam and adjusts itself.
2. Unsupervised learning — in this NN has to find some similarities. Let us say we give it a list of people and don’t say anything. GPT or the NN will try to find a common link or similarity between them and tell us. Something like all are engineers, or maybe all are from India or maybe even arrange them by their ages and reply. A use case is finding similarities. Say we feed the data on consumption in an area. The model could find patterns like what food is most trending in the area.
3. Reinforcement learning — in this, the NN is guided after each decision-making. Unlike supervised learning, where we give feedback via the final output, in RL, we give feedback after each decision it makes. A use case is when an NN is trying to learn to play chess. It doesn’t know whether the move is correct, but it gets points for each move. For a good move, it gets plus points; for a wrong one, it gets minus points. Based on this, it adjusts itself for maximum points.

Stay tuned for more such interesting things about the world of technology.