PaperlayerView exact revision on arXiv

Source-linked AI summary

ENCL: Enabling Efficient Neuromorphic Continual Learning with Local Learning Rule and Sub-Network Freezing Strategy

Rachmad Vidya Wicaksana Putra, Shuochen Zhou, Hassan Raza, Muhammad Shafique

arXiv:2610.31927v1cs.NEcs.AIcs.LG

TL;DR

Continual learning can update deployed neural networks with new run-time data, but retraining and existing NCL approaches can require substantial data and resources. ENCL combines local learning, task-specific sub-network freezing, and an efficient training policy. In task-incremental benchmarks, it reports accuracy comparable to data replay alongside lower memory, time, power, and energy costs.

  • Problem

    Continual learning for deployed systems must accommodate new operational data, while retraining can require unavailable old and new data and existing methods can impose high compute and memory costs.

    Source: Introduction · p. 1

  • Method

    ENCL combines a temporally- and spatially-local learning rule, task-specific sub-network freezing, and a training policy with fewer epochs and adjusted learning rate and timestep.

    Source: Introduction · p. 2

  • Results

    Across task-incremental benchmarks, ENCL achieves comparable accuracy to state-of-the-art data/experience replay while reducing memory cost by up to 88%, NCL training time by up to 51%, average power consumption by up to 40%, and energy consumption by up to 65%.

    Source: Conclusion · p. 7

  • Takeaways & Limitations

    The results position ENCL as a promising approach toward efficient and scalable continual learning for resource-constrained systems.

    Source: Conclusion · p. 7

    The high performance of the BPTT+ER approach comes at high cost from intensive backpropagation-based training and data/experience replay.

    Source: Maintaining High Accuracy across Tasks · p. 5

Abstract

from arXiv · show

Continual learning (CL) in neural networks is a promising learning approach for updating the knowledge of deployed systems based on new data encountered at run time, that have not been learned during the offline training phase. To enable their power/energy-efficient CL processing, neuromorphic continual learning (NCL) concept with sparse event-based operations have recently been introduced. However, state-of-the-art NCL methods typically employ the memory-hungry data/experience replay, coupled with costly global learning rule (i.e., backpropagation-based surrogate gradient). To address these limitations, we propose ENCL, a novel methodology that enables efficient NCL by employing a lightweight local learning rule and sub-network freezing strategy. ENCL leverages the temporally- and spatially-local learning rule to perform efficient credit assignment for training new tasks, alleviating the expensive cost of global learning rule. ENCL also employs a sub-network freezing strategy based on the given tasks to avoid the needs for network architecture expansion, hence further improving the efficiency and scalability of the NCL training process. To efficiently support this, we train each task with a small number of epochs as well as adjusted learning rate and timestep. Experiment results on task-incremental learning (TIL) scenario show that, our ENCL consistently achieves strong performance across different datasets as compared to state-of-the-art works. Specifically, ENCL achieves competitive accuracy comparable to that of the state-of-the-art data/experience replay, while reducing memory cost by up to 88%, NCL training time by up to 51%, average power consumption by up to 40%, and energy consumption by up to 65% across different benchmarking datasets (i.e., Split MNIST, Split Fashion MNIST, Split CIFAR-10, and Split CIFAR-100).

Introduction

ENCL targets continual learning under resource constraints, where retraining and existing NCL methods can demand substantial data, computation, memory, and energy. It combines local learning, sub-network freezing, and an efficient training policy; experiments report comparable accuracy to data replay with lower resource costs.

Introduction

  • Continual learning supports deployed systems learning from new operational data, while full retraining may be impractical when old and new task data are unavailable and can demand substantial time, computation, memory, and energy.Retraining may also cause catastrophic forgetting if not carefully performed.Source: Introduction · p. 1
  • Existing NCL methods are described as resource-intensive because they commonly rely on data replay and costly global learning rules, motivating alternatives for resource-constrained systems.The paper frames the goal as effective NCL with minimized computational and memory overheads.Source: Introduction · pp. 1–2
  • Comparable accuracy to state-of-the-art data replay accompanies reductions of up to 88% in memory cost, 51% in NCL training time, 40% in average power consumption, and 65% in energy consumption across four benchmarks.Evaluated in the task-incremental learning scenario on Split MNIST, Split Fashion MNIST, Split CIFAR-10, and Split CIFAR-100.Source: Introduction · p. 2
  • ENCL combines a temporally- and spatially-local learning rule for credit assignment, task-specific sub-networks that protect previously allocated parameters, and a training policy using fewer epochs with adjusted learning rate and timestep.The local rule is intended to alleviate the cost of global backpropagation-based surrogate-gradient learning, while sub-network freezing avoids network expansion.Source: Introduction · p. 2

Related Work

The related-work discussion contrasts global and local SNN learning rules and motivates TESS for supervised NCL. It also situates NCL in a pipeline with offline training, run-time task learning, and inference.

Related Work

  • Neuromorphic computing with SNNs uses sparse event-based data representation and operations, with reported hardware and software advances showing ultralow-power or energy processing across applications.Source: Related Work · p. 2
  • Global supervised learning rules such as surrogate-gradient methods can deliver high performance but incur substantial computational and memory overhead, whereas local rules are generally more efficient.The paper selects TESS, a temporally- and spatially-local supervised rule, for efficient NCL.Source: Related Work · p. 3
  • TESS addresses temporal credit assignment with eligibility traces encoding synaptic activity history and spatial credit assignment with learning signals derived from projected output spikes and targets.Weight updates combine the learning signals and eligibility traces.Source: Related Work · p. 3
  • The NCL pipeline comprises offline training at design time, NCL training at run time, and run-time inference; the paper evaluates task-incremental learning, where task identity is known and boundaries are clear.Each task typically contains multiple classes.Source: Related Work · p. 3, Figure 4 · p. 3

The ENCL Methodology

ENCL combines a fully local learning rule, task-specific sub-networks protected through freezing, and a tailored training policy for NCL. Its design assigns task pathways while reserving capacity for later tasks and adjusting training for new tasks.

The ENCL Methodology

Leveraging the Fully Local Learning Rule for NCL

  • ENCL integrates TESS at design time, trains known tasks offline, and reserves network space for NCL training.Task-specific sub-networks can handle tasks without expanding the network architecture.Source: Leveraging the Fully Local Learning… · pp. 3–4

The NCL Training Mechanism

  • Each task receives a masked sub-network; later tasks train only non-overlapping weights, while previously trained weights remain frozen.The mask occupies 1/K of network capacity, and saved task masks support task-specific evaluation.Source: The NCL Training Mechanism · p. 4
  • ENCL’s training policy uses fewer epochs, an adjusted learning rate, and a dataset-dependent timestep for new tasks.The adjusted learning rate supports more aggressive updates, while more complex datasets use a longer timestep.Source: The NCL Training Mechanism · pp. 4–5

Experiments

The experiments evaluate ENCL in task-incremental learning across four split image-classification benchmarks, measuring accuracy and training costs. Comparisons use non-NCL baselines and an experience-replay method on a shared network architecture.

Evaluation setup

  • Four benchmarks are evaluated: Split MNIST, Split Fashion MNIST, Split CIFAR-10, and Split CIFAR-100, with accuracy and NCL training costs measured.The evaluation runs on a single Nvidia RTX 4090 Ti GPU; outputs include training memory and time, and power and energy consumption.Source: Experiments · p. 5

NCL scenario

  • The 10-class benchmarks are split into five two-class tasks, while Split CIFAR-100 is split into ten ten-class tasks learned sequentially under TIL.Task 0 is learned offline; remaining tasks are learned in order during NCL.Source: Experiments · p. 5

Comparison partners

  • The comparison partners are BPTT-noNCL, TESS-noNCL, and BPTT+ER, with experience replay adapted from ANN to SNN.BPTT+ER uses the BPTT learning rule with data/experience replay.Source: Experiments · p. 5

Network architecture and settings

  • All methods use a VGG9 network with task-specific fully connected heads, while ENCL’s overlap threshold and timestep vary by benchmark.The three five-task benchmarks use Pthr=5% and T=5; Split CIFAR-100 uses Pthr=1% and T=9.Source: Experiments · p. 5

Results and Discussion

Across four TIL datasets, ENCL maintains accuracy broadly comparable to BPTT+ER while reducing training memory, time, power, and energy. The results cover accuracy preservation and the cost trade-offs of memory, training time, and power and energy consumption.

Maintaining High Accuracy across Tasks

  • ENCL preserves strong accuracy across tasks and datasets, finishing slightly above BPTT+ER on Split MNIST, Split Fashion MNIST, and Split CIFAR-100 but slightly below it on Split CIFAR-10.The reported final-task differences are about 0.5%, 0.1%, and 3% higher on the first three datasets, respectively.Source: Maintaining High Accuracy across Tasks · pp. 5–6, Figure 5 · p. 6

Minimizing Memory Cost during NCL Training

Minimizing NCL Training Time

Optimizing Power and Energy Consumption during NCL Training Process

Conclusion

ENCL combines a fully local learning rule with sub-network freezing to support efficient NCL without network expansion. In TIL experiments, it achieves accuracy comparable to replay while reducing memory cost, training time, power, and energy across four benchmark datasets.

Conclusion

  • ENCL combines a fully local learning rule and sub-network freezing to avoid network expansion while achieving comparable accuracy to BPTT+ER and reducing memory, training time, power, and energy costs across four TIL datasets.The reported maximum reductions are 88% in memory cost, 51% in NCL training time, 40% in average power consumption, and 65% in energy consumption.Source: Conclusion · p. 7

Related papers

Loading 2610.31927v1…