Models within a model: in-context learning pushes boundaries of LLM

The concept itself is quite simple and easy to understand. However, due to the high level of abstraction, it is difficult to look under the hood, i.e. to understand exactly how a neural network arrives at its outputs because there are many neurons and (as with the brains of living organisms) it is not easy to trace how any given neuron contributes to the outcome precisely.

Surprising abilities of large language models

A very interesting situation arises when we look at some of the surprising abilities of these modern neural networks. Researchers have recently noticed a curious phenomenon: so-called in-context learning. It is a phenomenon that experts do not yet fully understand – the LLM can be effectively applied to perform a new surprise task without being trained to do so (i.e., to change the model’s weights directly; see box 2). The LLM can adjust its behaviour during so-called model inference, where we only give the model inputs and compute the output by forward propagating the signal through the layers of neurons.

In this case, we give the LLM an input sentence with the expected answer at the end. It is fascinating that it can deduce the relationships and respond with the correct output.

Source: How does in-context learning work? A framework for understanding the differences from traditional supervised learning, The Stanford AI Lab Blog

As another example, consider the popular Midjourney image generator. It was trained on a huge pile of images from all over the Internet, which it then learned from; at first it was given the task of painting simple things and gradually worked its way up to much more sophisticated ones.

The workings of the process within the neural network are as follows. We find a linear function that can paint a particular thing, for example, an arc, and it then uses that to paint a leaf and so on. This function did not exist before the model was trained – the model itself created it.

Then when you want something else after you have pre-trained the model in this way, it makes it easier for it. Suddenly it doesn’t need a vast number of input examples, but it can cope with a few. For instance, if given the task of drawing a human face, it will use an existing linear model, but instead of creating a leaf, it will create a piece of an ear.

The algorithm had to change to adapt to the new task to do this. That happens somewhere in that middle part of the abstraction, the part we can’t see, that connects the linear functions into a usable output, in this case, the generated image.

When you want something else after the pre-trained model, it makes it easier for it. Suddenly it doesn’t need a vast number of input examples, it can cope with a few.
The algorithm had to change, and that happens somewhere in that middle part of the abstraction, the part we can’t see.

A machine learning model would have to be retrained with new data by default. It is a fundamental principle of algorithms and models work that way by default. As the model is trained, it updates existing parameters as it processes new information. However, with in-context learning, there is no predetermined change in parameters. Thus, the model – at first glance – learns the new task without changing its weights.

Research by MIT, Google Research, and Stanford University scientists shows that these massive neural networks deep down contain smaller, simpler linear models that the model acquired during its initial training. The language model then uses a simple algorithm to train this smaller, linear model to perform the task by using information already in the large model itself.

It is, actually, a model within a model.

Smarter models for less money

Let’s take GPT as an example. The model has hundreds of billions of parameters and was trained on a huge amount of text from the Internet. So if someone presents a new task to the model, it is very likely that the model will be able to respond appropriately. In fact, it has very likely already encountered something similar during training. The model does not learn to respond to new tasks in isolation but repeats patterns it has already seen during its training. In the output, it uses a previously trained smaller model hidden inside the overall model.

But why have researchers yet to learn exactly how LLMs work? Of course, they know in a cursory way, but what happens in neural networks between input and output is partly hidden from the experts. The different layers of interconnected neurons simply “work” somehow – the level of abstraction is already very high indeed. It is not magic, but it is a fascinating phenomenon confirming the advanced nature of artificial neural networks; it is a surprisingly intelligent process we have yet to understand in depth. It makes neural networks and large language models all the more interesting.

In-context learning is potentially one of the most important functions of neural networks. If understood thoroughly, it could be applied so that models do not need to be retrained for new tasks, which is otherwise quite expensive.

The time is rapidly approaching when AI will become even more prominent in everyday human life. Our knowledge, understanding, and, likely, the nature of work and productivity will change. General artificial intelligence, a concept that has fascinated and terrified science fiction authors for generations, is steadily approaching. The issue is time and the limits humans set for themselves: The better we understand AI, the safer, better and more sustainably we can use the latest technology.