Neural networks adapt to new tasks using local memory updates
Online Task Adaptation via Self-Organisation
Machine LearningNeural and Evolutionary Computing
Summary
Neural networks usually learn by adjusting their many settings through math called gradients. This paper shows that instead, networks can be trained first to organize their own learning through small local changes in memory without needing to calculate gradients again. The researchers used a special network made of tiny interacting cells with memory that updates itself when it sees errors. After training, the model can quickly improve on new tasks by just changing these local memories once, without changing the main settings. This approach works well even when the new tasks are quite different from what the network saw before.
What this means in practice
- •For machine learning engineers: Develop adaptive systems that improve on new classification tasks without expensive gradient computation.
- •For edge device developers: Enable devices that update their models locally and quickly from few examples without accessing central servers or retraining.$Commercial implications: This paper enables lightweight on-device personalization for AI products, reducing need for cloud computation and data transfer.
Authors
Krsto Proroković
Abstract
Neural networks are typically adapted by computing gradients and updating model parameters. We investigate whether task-specific adaptation can instead emerge from a meta-learned self-organising process that requires no gradients at adaptation time. We instantiate this idea with a Neural Cellular Automaton in which locally interacting recurrent cells maintain both a recurrent state and a fast associative memory. During meta-training, backpropagation is used to learn the recurrent dynamics together with how the memory is read and written. Once training is complete, the slow model parameters remain fixed, and online adaptation occurs only through cellwise memory updates driven by local prediction errors and a delta rule. We evaluate whether the learned mechanism can adapt to semantically distinct held-out classification tasks. A single pass over the support data produces substantial improvements in held-out performance without gradient computation or parameter updates during adaptation, and the mechanism remains effective across large changes in the number of examples processed jointly. These results show that task-specific adaptation can be achieved through explicit fast-memory updates while keeping the slow model parameters fixed.