Scientists believe they might have found a way to overcome “model collapse”, a phenomenon that threatens the training of AI as we know it.
To become better, artificial intelligence systems such as those used in ChatGPT must be trained with more and more real data. But much of that data is taken from writing on the internet, which itself is often produced using such models.
The amount of real data is rapidly reducing – and is predicted to run out as early as this year. And data that is produced by other AI systems could rapidly lead to “model collapse”.
That refers to a phenomenon where AI systems engage in “data cannibalism”, where they are trained on their own outputs, and come rapidly less useful and more prone to dangerous falsehoods.
But researchers have suggested that using just one datapoint from the outside world can prevent the problem. They did so using a set of models called “Exponential Families”, a set of statistical models.
The work showed that training systems only on data it has produced will always lead to model collapse. But incorporating one datapoint from outside of the model – such as previously acquired knowledge – prevents the effect.
That happens even when the amount machine-generated data is infinitely large, the study showed.
“Previous work undertaken on model collapse primarily looks at large, complicated LLMs, where it’s not clear how these models work and if results are repeatable – it is why you get unexplained hallucinations, where you can’t explain why an AI has generated a wrong answer,” said Yasser Roudi, Professor of Disordered Systems at King’s College London.
“By focusing on a simple model, we can establish why adding just one data point prevents them from generating gibberish from an objective, statistical standpoint.
The researchers behind the study pointed out that such collapse is not limited to chatbots – but could affect vital infrastructure such as cars, too.
“From this foundation, we can establish principles that will be vital in future AI construction,” said Professor Roudi. “As larger models are deployed in areas touching our lives, from ChatGPT to self-driving cars, and synthetic data takes on a larger share of AI training, computer scientists will have the tools to prevent this potentially disastrous scenario.”
The work is reported in a paper published in the journal Physical Review Letters.
