The premise of large language models (LLMs) is that they continue to improve as they are scaled up. This is in stark contrast to the conventional wisdom of statistical learning theory, which posits that smaller models should outperform larger ones. Just a few years ago, the idea of massive LLMs outperforming smaller ones would have been unthinkable. The technology that powers LLMs, namely auto-regressive transformer neural networks, has been around since 2017, yet the focus was not on creating larger models. The reason for this lies in the concept of overfitting.
5 years ago, nobody would have guessed that scaling up LLMs would as successful as they are. This belief, in part, was due to the fact that all known statistical learning theory predicted that massively oversized models should overfit, and hence perform worse than smaller models.
Yet the undeniable fact is that modern LLMs do possess models of the world that allow them to generalize beyond their training data.
This video is from Algorithmic Simplicity.
In machine learning, models learn from a set of examples. Ideally, the model should learn the underlying mechanism or rule behind the data, allowing it to generalize to new situations. However, a model may instead memorize the training data, a phenomenon known as overfitting. This means the model cannot effectively deal with new inputs. One way to prevent overfitting is to reduce the size of the model, forcing it to learn the underlying mechanism rather than memorizing the data.
This understanding led to the belief that training massive LLMs was not a good idea, as they would simply overfit and regurgitate memorized information. AI researchers focused on developing better learning algorithms and more effective neural network architectures. However, a 2019 study demonstrated that as neural networks were scaled beyond the point of overfitting, their performance actually improved, a phenomenon known as double descent. This surprising result seemed to defy statistical learning theory and suggested that simply increasing model size could improve AI performance.
Around the same time, another study found that most of the weights in a large neural network are essentially useless, and a much smaller network could achieve the same performance. The study proposed that training a large network was equivalent to training all of its smaller subnetworks independently and choosing the best one. This concept is known as the lottery ticket hypothesis. The idea is that, with enough “lottery tickets” or subnetworks, a winning network with a good initialization will eventually emerge, allowing even smaller subnetworks to perform well.
This counterintuitive result means that the larger the neural network, the smaller the winning subnetwork will be, resulting in a simpler learned model. This explains why larger neural networks generalize better, and aligns with statistical learning theory’s preference for simpler models.
The lottery ticket hypothesis also suggests that the pruning strategy used in the study is not optimal, and the true size of the minimal subnetwork is likely much smaller than identified. This raises questions about the future of AI and the potential for human-level intelligence. The smallest program that generates all the text on the internet would likely be a physics engine that simulates human brain interactions, suggesting that a perfect learning algorithm could, in principle, implement such a program.
While modern neural networks fall far short of this ideal, they do show some ability to generalize beyond their training data. Scaling up LLMs could bring them closer to this ideal learner. However, the rate of improvement with increased scale may be too slow to achieve human-level intelligence. Nonetheless, the potential for LLMs to reach human-level intelligence through better learning algorithms remains an intriguing possibility. Regardless of the ultimate outcome, the journey of exploring and understanding the capabilities of LLMs continues to be a fascinating exploration of the frontiers of artificial intelligence.
