Source-linked AI summary
Online Task Adaptation via Self-Organisation
Krsto Proroković
TL;DR
Neural-network adaptation typically requires a separate gradient-based backward computation, and the paper asks whether task-specific learning can instead occur without gradients during adaptation. It meta-learns recurrent cells with fast memory that adapts from local prediction errors while slow parameters stay fixed. One support-set pass raises mean held-out accuracy from 20.0% to 48.2%, compared with 54.4% for a corresponding scratch-trained baseline.
Problem
The paper asks whether a model can adapt from prediction errors without backpropagating gradients through its computation during adaptation.
Source: §1 · p. 1Method
A meta-trained Neural Cellular Automaton uses local prediction errors and a delta rule to update cellwise fast memory, leaving slow parameters fixed during adaptation.
Source: §1 · p. 1, §2.3 · pp. 2–3Results
A single support-set pass raises mean held-out accuracy from 20.0% to 48.2%, compared with 54.4% for a corresponding baseline trained from scratch.
Source: §1 · p. 1Takeaways & Limitations
The results show that task-specific adaptation can be achieved through explicit fast-memory updates without gradient computation during adaptation.
Source: §1 · p. 1The evaluation uses superclass tasks rather than FC100's mixed-superclass task sampling, so results are not directly comparable to published FC100 results.
Source: §4.1 · p. 4
Abstract
from arXiv · showhide
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.
1 Introduction
The section asks whether task-specific learning can occur without gradient propagation during adaptation and introduces a Neural Cellular Automaton using locally updated fast memory. A single support-set pass improves held-out accuracy while slow parameters remain fixed.
1 Introduction
- The proposed Neural Cellular Automaton uses locally interacting recurrent cells with fast associative memory, updated from local prediction errors using the delta rule.During inference, cells exchange information and read from memory to form predictions.Source: §1 · p. 1
- The authors frame task-specific behavior as self-organisation emerging from recurrent interactions and cellwise memory updates.The distinction from standard adaptation is that gradients are not propagated through the prediction computation.Source: §1 · p. 1
- Meta-training learns recurrent dynamics and memory read/write behavior, while online adaptation changes only fast memory and computes no gradients.Slow model parameters remain fixed after training.Source: §1 · p. 1
- A single support-set pass raises mean held-out accuracy from 20.0% to 48.2%, versus 54.4% for a corresponding baseline trained from scratch.The baseline uses backpropagation for task-specific training.Source: §1 · p. 1
2 Method
The method uses interacting recurrent cells with fast associative memory, which stays fixed during inference and is updated from local prediction errors. Meta-training learns the dynamics and memory operations, while task processing uses sequential mini-batches to adapt the memory and train the slow parameters.
2 Method
- The method is implemented with an NCA but could also use other architectures that iteratively update interacting units with shared parameters.The stated examples include Universal Transformers, looped Transformers, and recurrent graph neural networks.Source: §2 · p. 1
2.1 Inference
- Each cell maintains recurrent state and fast associative memory; during inference, memory stays fixed as recurrent state evolves, while cells communicate through neighboring states.The method is presented for image classification but is described as adaptable to other input structures and supervised adaptation tasks with target outputs.Source: §2.1 · pp. 1–2, §2 · p. 1
2.2 Adaptation
- Cells compare their outputs with targets, then use learned write keys and values to update memory by the delta rule, with batch updates scaled to match meta-training.Each update moves the value retrieved at a write key toward the computed write value; batches larger than the meta-training batch size are processed sequentially.Source: §2.2 · p. 2
2.3 Meta-training
- Meta-training processes each task as sequential labelled mini-batches: an initial batch adapts memory without contributing to loss, while later batches provide online cross-entropy objectives.Slow parameters are optimized by backpropagation through windows of batches, with memory carried across window boundaries but gradients not propagated through preceding windows.Source: §2.3 · pp. 2–3
3 Related work
Related work spans gradient-based parameter adaptation, learned plasticity, and fast-weight memories. This method is distinguished by its explicit cellwise memory and prescribed update rule.
3 Related work
- MAML and its variants adapt parameters with gradient-based optimization, while learned optimizers meta-learn update rules and in-context methods condition on task demonstrations.The paper contrasts these approaches with incrementally updated explicit fast associative memory as task context.Source: §3 · p. 3
- Differentiable Plasticity, Backpropamine, and related work meta-learn plasticity mechanisms; Variable Shared Meta Learning can adapt without explicit gradient computation.This method shares the goal of gradient-free meta-test adaptation while separating fixed slow parameters from cellwise fast memory updated by a prescribed local rule.Source: §3 · p. 3
- Fast-weight models, Memory-Augmented Neural Networks, Meta Networks, and SRWM also separate slowly learned parameters from fast task-dependent adaptation.Their approaches include temporary fast weights, separate memory, gradient-derived fast weights, and runtime self-modification, respectively.Source: §3 · p. 3
- Unlike SRWM, this method uses spatially distributed cell memories with error-determined delta-rule writes and can aggregate parallel batch updates.SRWM learns write strength and updates sequentially; this method also uses the same online stream for memory updates and meta-training loss, rather than separate examples.Source: §3 · p. 3
4 Experiments
The experiments test fast-memory adaptation on semantically shifted CIFAR-100 superclass tasks and compare it with training from scratch. One support pass yields substantial held-out accuracy gains, while results remain robust to support-batch size and vary with truncation window during repeated adaptation.
4.1 Experimental setup
- Tasks are five-class CIFAR-100 superclasses split into 12 meta-training, 4 validation, and 4 test superclasses, creating semantically distinct held-out tasks.At meta-test time, adaptation uses 2500 training-split images and evaluation uses 500 test-split images, with memory frozen during evaluation.Source: §4.1 · p. 4
4.2 Results
- After one support pass, mean held-out accuracy rises from 20.0% to 48.2%, versus 54.4% for scratch training, recovering approximately 82% of the chance-to-scratch gap.The adapted model sees each support image once and changes only its fast memory; the scratch baseline uses gradient-based training over tens of passes.Source: Table 2 · p. 5, §4.2 · p. 5
- Across support batches of 1, 16, 32, 64, and 128, mean accuracy is 48.3%, 48.3%, 48.4%, 48.1%, and 48.2%, respectively.These groupings change memory writes from 2500 to 20; the passage attributes the small practical effect to fixed per-example scaling despite changed correction recomputation frequency.Source: §4.2 · p. 5
- On people, fast-memory adaptation reaches 33.1% accuracy versus 36.2% from scratch; the authors suggest task difficulty partly reflects the 32 × 32 inputs and architecture.The task distinguishes baby, boy, girl, man, and woman; the scratch baseline uses 2250 labelled examples.Source: §4.2 · p. 5
4.3 Effect of truncation window across adaptation passes
- For L=4 and L=8, accuracy peaks on pass two and then loses fewer than three points over eighteen more passes, while L=22 reaches 29.4% by pass twenty.L=22 spans all loss-contributing batches within a meta-training pass; full backpropagation through that pass yields the least stable repeated memory updates in this setting.Source: §4.3 · p. 5, Figure 1 · p. 6