Nested Learning - a new paradigm in machine learning

Google Research published a paper at NeurIPS '25 : "Nested Learning: The Illusion of Deep Learning Architectures"

It is a new approach to look at machine learning problems: architecture & optimization algorithms together

In NL, training models is not treated as one opaque, atomic process. But as a system of interconnected, multi-level learning problems that are optimized in parallel.

The goal is to solve one of the core challenges of modern LLMs: continual learning (the ability for a model to acquire new skills over time without forgetting old ones.)

Authors argue: the model's architecture and the optimization algorithm are fundamentally the same concepts; they are just different "levels" of optimization, each with its own internal flow of information and update rate

We'll walk through:
- how the paper unifies model architecture and optimizers
- architectural decomposition of a classical learning problem solved in the framework of Nested Learning
- how optimizer algorithms fit into the NL framework (backrpop as associative memory)
- the concept of continuum, multi-timescale memory
- Hope - an architectural backbone of NL

=== ENTRY DETAILS ===

- QR code with entry information will beavailable soon, in the "Photos" section of this event page.
- Gate closes at 18:15 - no late entries.