Google Research has developed a new machine learning technique called Nested Learning that is designed to address a flaw in AI called catastrophic forgetting in continual learning. Catastrophic forgetting, simply put, is when an AI is updated with new information or given a new skill to learn and wipes the slate, becoming worse at the things it knew previously.

The new Nested Learning method that has been developed at Google Research helps to solve the problem of catastrophic forgetting by taking inspiration from the human brain. Google called it a robust foundation for bridging the gap between current LLMs and their forgetful nature and the human brain.

Nested Learning involved a paradigm shift in thinking about an AI model’s architecture and optimization algorithm. Traditionally, AI developers have viewed the model’s architecture and optimization algorithm as two separate entities; however, with Nested Learning, the idea is to view them as one entity.

Nested Learning treats AI models as a series of smaller, interconnected, nested optimization problems. Each of these sub-problems are allowed to learn and update its knowledge at a different specific pace in a technique called multi-time-scale updates. This simulates the function of neuroplasticity in the human brain, where parts of your brain adapt to new experiences without scrubbing old memories. So, instead of the whole model updating uniformly and wiping old knowledge, nested learning allows it to have a layered, dynamic system that updates without erasing old knowledge.

Google Research applied its proposed principles to design a proof-of-concept model called Hope. It’s a self-modifying recurrent architecture that can optimize its own memory. It uses Continuum Memory Systems, which treats memory not as simple short-term and long-term buckets, but as a spectrum of layered memory modules, each updating at its own frequency. This allows the model to manage and retain more data over time in a richer and more organized way.

Hope was able to consistently outperform current state-of-the-art models in long-context memory challenges, such as Needle-In-Haystack tasks, where it has to recall a specific, small detail buried deep in a large document. The model was also more accurate and efficient in general language modeling.

We will hopefully see these improvements in upcoming versions of Google Gemini. Unfortunately, Google did not share a timeline of when we can expect to see it land in its flagship AI model.

Source: Google Research