Introducing Nested Learning: A new ML paradigm for continual learning

Research and Theory
1

This new Google research paper (final year PhD student) feels very thousand brainy to my spider sense. Islands of knowledge, separated by time based propagation feedback loops between independent learners. It is even explicitly inspired by neuroscience (section 1.1 of the paper not the overview.) when I look at the some of the details of the hope architecture it is reminiscent of an early Numenta paper/code trying to make continual learning by making a pyramidal neuron.

2

Hey @ricpruss, there’s an existing topic about it: "Introducing Nested Learning: A new ML paradigm for continual learning" by Google

I was critical of it at the time, but a follow-up preprint paper by the same authors paints a clearer and much more interesting picture: [2606.03979] Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories

Essentially, during “sleep”, an RL-based student-teacher framework blends data from the context window into frequency-layer parameter updates for short-term memory integration, while avoiding catastrophic forgetting. Although, they talk very little about performance and scalability, and well, it’s still… RL, with the downsides that come with it.

3

Hi @ricpruss, thanks for sharing! And thanks @AgentRev for pointing to the existing topic and also the follow-up paper.

My two-cents after skimming the papers (especially the later “Language Models Need Sleep”): It’s nice to see papers that are neuroscience-inspired, though I’d still say that “inspired” is doing a bit of the heavy lifting. The actual methods implemented in the 2nd paper like distillation and augmenting with synthetic data doesn’t seem to have basis to neural mechanisms except the analogies to sleep and dreaming, respectively. The performance improvement based on Table 1 (comparing Sleep vs. OPSD + Parameter Expansion) also seems pretty small - but I’m not an expert to know if these are significant or not.

Regarding the Nested Learning paper, we do talk quite a bit about Hippocamus and Fast Learning in our research meetings, e.g. most recently 2026/04 - Focus Week Recap on Temporary Models in the Hippocampus if you are curious in checking out!

Appreciate the share. :slight_smile:

4

Oh wow, thanks to that link on the hippocampus and I think I should watch your recap meetings that was quite the education!

1 Like