Source-linked AI summary
Fast On-Device Adaptation for Spiking Neural Networks via Online-Within-Online Meta-Learning
Bleema Rosenfeld, Bipin Rajendran, Osvaldo Simeone
TL;DR
Personalized edge applications need SNNs that adapt to new tasks from limited streaming data, while existing SNN meta-learning methods rely on offline training and backpropagation. OWOML-SNN addresses this gap with nested online local updates for probabilistic SNNs. It improves few-shot and overall adaptation accuracy over conventional and joint-training benchmarks while supporting continual online task adaptation.
Problem
Personalized edge applications require adaptation to individual data, but existing SNN meta-learning methods operate offline and rely on backpropagation.
Source: §I · pp. 1–2Method
OWOML-SNN uses online-within-online meta-learning with probabilistic GLM SNNs, local nested updates, and parallel task-specific models initialized from shared weights.
Source: §I · p. 2, §III · p. 3, §VI · p. 4Results
OWOML-SNN achieves about 18% larger accuracy improvement than joint training after five examples per class and 15% higher average overall accuracy than two benchmarks.
Source: §VII · p. 5Takeaways & Limitations
The approach supports fast online adaptation to sequentially observed tasks without backpropagation, making it a candidate for on-chip adaptation.
Source: §VII · p. 6, §VIII · p. 6
Abstract
from arXiv · showhide
Spiking Neural Networks (SNNs) have recently gained popularity as machine learning models for on-device edge intelligence for applications such as mobile healthcare management and natural language processing due to their low power profile. In such highly personalized use cases, it is important for the model to be able to adapt to the unique features of an individual with only a minimal amount of training data. Meta-learning has been proposed as a way to train models that are geared towards quick adaptation to new tasks. The few existing meta-learning solutions for SNNs operate offline and require some form of backpropagation that is incompatible with the current neuromorphic edge-devices. In this paper, we propose an online-within-online meta-learning rule for SNNs termed OWOML-SNN, that enables lifelong learning on a stream of tasks, and relies on local, backprop-free, nested updates.
I. INTRODUCTION
The paper motivates online adaptation for personalized, low-power edge intelligence and proposes OWOML-SNN, an online-within-online meta-learning rule for SNNs that avoids offline pretraining and backpropagation.
Context and motivation
- Personalized mobile applications need models that adapt after deployment to users’ and local environments’ unique data distributions.The paper highlights speech impairment or accent as an example where fixed offline-trained models can produce inaccurate or unfair results.Source: §I · p. 1
- Meta-learning replaces a fixed universal model with an adaptation procedure that adjusts models using limited local data.Source: §I · p. 1
Online vs offline meta-learning
- Online meta-learning processes streaming data sequentially, unlike offline procedures that assume fully accessible static datasets.Source: §I · p. 2
Main contributions
- SNNs combine sparse binary activations with low-power neuromorphic hardware, making them suitable for on-device adaptation.Neuromorphic chips restrict learning to local rules, potentially with global learning signals.Source: §I · p. 2
- OWOML-SNN performs online-within-online meta-learning for SNNs without offline pre-deployment training or backpropagation.Its nested local rule is derived using probabilistic GLMs and variational inference.Source: §I · p. 2
II. RELATED WORK
Prior meta-learning work includes several online and offline loop configurations, but SNN approaches generally train offline and use surrogate-gradient backpropagation; this paper instead uses probabilistic GLM neurons for direct local-rule derivation.
Online vs offline meta-learning
- The proposed setting combines streaming outer-loop task data with streaming within-task data, unlike related methods using batch inner-loop learning.The paper contrasts its dynamic probabilistic SNN with a static deterministic linear model under a convex loss.Source: §II · p. 2
SNN meta-learning
- Existing SNN meta-learning methods support online local within-task learning but perform meta-training offline or before deployment.Source: §II · p. 2
- Prior SNN methods use deterministic LIF neurons and surrogate gradients, whereas this work adopts probabilistic GLM neurons for local maximum-likelihood rules.Source: §II · p. 2
III. ONLINE-WITHIN-ONLINE META-LEARNING
Online-within-online meta-learning updates a task-specific model from streaming current-task data while concurrently improving a shared initialization using sampled previous tasks.
Online-within-online formulation
- The method uses a shared hyperparameter θ as the initialization for efficient within-task learning across a family of statistically related tasks.Source: §III · p. 2
Data buffers
- Streaming batches accumulate in a task-data buffer, which is added to a meta-data buffer after the current task is processed.Source: §III · p. 2
Concurrent updates
- At each within-task step, the current task model is updated from its accumulated data while the shared initialization is concurrently updated from previous tasks.Source: §III · pp. 2–3
Meta-update
- The meta-update samples N previous tasks, trains parallel task-specific models from θ, and applies a gradient update with meta-learning rate µ.Source: §III · p. 3
Meta-gradient
- The exact meta-gradient requires second-order derivatives, so the method uses the first-order REPTILE approximation.Source: §III · p. 3
IV. SNN MODEL
The model is a recurrent probabilistic GLM-based SNN that processes binary spikes through visible and hidden neurons, with trainable temporal filters and locally estimable learning signals.
SNN architecture
- The recurrent SNN processes binary spikes over time using visible neurons driven by data and hidden neurons supporting the network’s operation.Connections may exist within and across the visible and hidden neuron sets.Source: §IV · p. 3
Neuron dynamics
- Each membrane potential integrates pre-synaptic and feedback spikes through trainable temporal filters plus a trainable bias.The filters operate over windows of previous spikes and use raised-cosine basis functions.Source: §IV · p. 3
Probabilistic neuron
- In the GLM neuron, the probability of a spike is the sigmoid of its membrane potential and depends on previously integrated spikes.Source: §IV · p. 3
V. PER-TASK LEARNING
The per-task learner performs online variational training of a probabilistic SNN by processing task examples sequentially and updating parameters with local gradient signals. Eligibility traces accumulate these updates, while hidden-neuron learning uses a global error signal in a three-factor rule.
Per-task update
- The per-task Update function concatenates task examples into one stream and iteratively maximizes the observed-spike log-likelihood from initialization θ.The objective is approached through a lower bound on the log-likelihood.Source: §V · p. 4
- Monte Carlo samples of hidden spikes estimate the expected log-likelihood and enable recurrent membrane-potential updates at every discrete time step.The model parameters φ are initialized by θ and updated from the resulting gradient.Source: §V · p. 4
Local gradients
- Visible-neuron derivatives combine a post-synaptic error term with a pre-synaptic filtered spike-history term defined by basis functions.The pre-synaptic history spans a τw-step window.Source: §V · p. 4
- Hidden-neuron gradients use REINFORCE, with the learning signal accumulated alongside other derivatives in exponentially decaying eligibility traces.The accumulated derivatives are applied after Δs discrete time steps.Source: §V · p. 4
Local learning rule
- The resulting three-factor hidden-neuron update is local except for a global error signal and includes a sparsity-inducing regularizer.The global signal indicates how well hidden neurons support the desired output.Source: §V · p. 4
VI. SNN META-LEARNING
OWOML-SNN performs online-within-online meta-learning by maintaining an inference SNN and additional models that support meta-updates as task data arrives. Its goal is rapid adaptation to new tasks using fewer labeled examples.
Algorithm
- OWOML-SNN instantiates N+1 SNNs from the current hyperparameter initialization, using one for inference and N for the meta-update.The hyperparameters are represented by the meta-learner’s weights θ(t,i).Source: §VI · p. 4
Within-task learning
- The inference SNN updates online on a growing buffer of current-task data using Update(θ(t,i), D(t,i)).New batches are added at each within-task time step.Source: §VI · p. 5
Meta-update
- The meta-update trains the remaining SNNs online on datasets sampled from the meta-data buffer of previously seen tasks.Each sampled dataset contains M examples from a prior task.Source: §VI · p. 5
Evaluation objective
- Online meta-learning succeeds when the inference SNN reaches satisfactory accuracy on held-out current-task data using fewer adaptation examples.This criterion evaluates both accuracy and adaptation speed.Source: §VI · p. 5
VII. EXPERIMENTS
Experiments compare OWOML-SNN with conventional per-task training and joint training on Omniglot and MNIST-DVS task families. The results show stronger few-shot adaptation and continued improvement as online meta-training proceeds.
Baselines
- The comparisons use conventional training with random reinitialization per task and joint training across examples from the task buffer as reference baselines.These baselines disable or replace the meta-update function, respectively.Source: §VII · p. 5
Omniglot
- OWOML-SNN is evaluated on Omniglot 2-way tasks containing characters that may come from different alphabets, with 14 examples per class and six held out for testing.Images are downsized to 26×26 pixels and rate encoded over T = 80.Source: §VII · p. 5
- After 5 examples per class, OWOML-SNN’s accuracy increase over conventional training is about 18% larger than joint training, and later accuracy is 15% higher on average than both benchmarks.The Omniglot evaluation tests six unseen tasks after t = 15 meta-time steps.Source: §VII · p. 5
MNIST-DVS
- On MNIST-DVS, OWOML-SNN enables fast online adaptation that continuously improves as meta-time steps and meta-updates increase, requiring fewer examples to achieve within-task accuracy targets.Results are shown after the full current-task dataset D(t,i) has been processed, with M = 4, N = 5, and Δs = 20.Source: Figure 4 · p. 5, §VII · p. 6
VIII. CONCLUSIONS
The paper presents OWOML-SNN as a local, backpropagation-free online meta-learning method for sequential task adaptation. Its first-order meta-gradient improves efficiency and locality but generally reduces adaptation capacity.
Conclusion
- OWOML-SNN demonstrates performance benefits when adapting probabilistic SNNs to new tasks observed sequentially online.The method builds on online variational learning.Source: §VIII · p. 6
- A first-order meta-gradient yields a local meta-learning rule based on predictive coding without backpropagation of gradients.This makes the approach a candidate for fast on-chip adaptation, particularly with streaming inputs.Source: §VIII · p. 6
Limitations and future work
- The first-order approximation generally reduces the capacity to adapt to new tasks, motivating more accurate local meta-gradient approximations and convergence analysis.These are identified as future research directions.Source: §VIII · p. 6