Source-linked AI summary
RobustNCL: Robustifying Neuromorphic Continual Learning under Transient Faults with Lightweight Detect-and-Cure Strategy
Rachmad Vidya Wicaksana Putra, Shuochen Zhou, Muhammad Shafique
TL;DR
Transient DRAM faults threaten NCL accuracy, while NCL reliability under these faults and corresponding mitigation techniques have not been studied. RobustNCL detects anomalous weights during DRAM access and cures them with selected replacement values. Across NCL methods and datasets, it maintains accuracy close to fault-free models and improves on unmitigated baselines.
Problem
Transient faults in DRAM-stored weights can undermine NCL accuracy, but NCL reliability under these faults and corresponding mitigation techniques have not been studied.
Source: §1 · p. 2Method
Results
The best RobustNCL variant maintains accuracy within -0.5% of fault-free models and significantly improves accuracy over unmitigated baselines across NCL methods and datasets.
Source: §6 · p. 9Takeaways & Limitations
RobustNCL offers a potential approach to reliable, energy-efficient NCL, with negligible latency and energy overheads reported across evaluated settings.
Source: §6 · p. 9The evaluation uses FP32 weights and a single DDR3-1600 2Gb DRAM chip to store the investigated model weights.
Source: §3.2 · p. 5
Abstract
from arXiv · showhide
In neural networks, continual learning (CL) is a prominent approach for enabling deployed systems to adapt to dynamic operational environments by effectively learning from new unseen data at runtime. Recently, the neuromorphic continual learning (NCL) concept has emerged as an energy-efficient alternative to CL by leveraging sparse event-based operations. However, NCL performance can be undermined by transient faults in the hardware platforms, which may come from high-energy particle strikes at any time. Currently, NCL reliability under transient faults has not been studied, and hence the corresponding fault-tolerance techniques remain unexplored. To address this, we propose a novel RobustNCL framework to improve NCL reliability in the presence of transient faults in the DRAM-based off-chip memory of hardware platforms through a lightweight detect-and-cure strategy. Specifically, it first monitors the presence of faults during weight data access from off-chip memory, and then cures the accessed weights if anomalous signatures are detected. It provides three curing variants: replacement with zero, min/max value, and most-probable value from the original/clean weight distribution. To support this, the weight curing values are stored in the radiation-hardened memory. Experimental results show that, in the presence of high fault rates (e.g., 5x10^{-4} for memory replay), RobustNCL maintains high accuracy close to the accuracy of fault-free models (i.e., within -0.5% accuracy for the best RobustNCL variant) and significantly improves accuracy from the baseline without fault mitigation across different NCL methods and datasets (i.e., Split MNIST, Split Fashion-MNIST, Split CIFAR-10, and Split CIFAR-100), while incurring negligible latency and energy overheads. These results highlight that our RobustNCL framework is a promising approach for enabling reliable and energy-efficient NCL systems.
1 INTRODUCTION
Continual learning helps deployed systems adapt to new tasks, and neuromorphic continual learning offers an energy-efficient approach, but transient DRAM faults threaten its accuracy. RobustNCL addresses the understudied reliability problem with a lightweight detect-and-cure strategy and a fault model and evaluation for NCL.
1 INTRODUCTION
- RobustNCL targets transient faults in DRAM-stored weights, a threat to NCL accuracy whose reliability implications and mitigation techniques had not been studied.The paper identifies high-energy particle strikes as one source of transient faults.Source: §1 · p. 2
- The work introduces a transient-fault model and experimentally analyzes NCL fault tolerance across fault rates and NCL methods.Faults are applied to DRAM at different possible points in time.Source: §1 · p. 2
- RobustNCL detects anomalous DRAM-accessed weights and replaces them with curing values selected to support safe spiking activity.Curing values are informed by observations during NCL training, accounting for possible changes in model state.Source: §1 · p. 3
- The framework evaluates combinations of fault rates, datasets, and the Subnet and memory replay NCL methods.Figure 2 depicts fault exposure, detect-and-cure with curing-value selection, and NCL training; variants differ by how faulty weights are replaced.Source: §1 · p. 3, Figure 2 · p. 3
2 RELATED WORK
The paper situates its work in neuromorphic continual learning, which combines event-driven spiking computation with continual learning, and considers two NCL methods. It also describes the task-incremental setting and the DRAM hierarchy relevant to its study.
2 RELATED WORK
- NCL combines event-driven spiking neural network computation with continual learning and proceeds through offline training, runtime NCL training, and runtime inference.The paper studies task-incremental learning, where tasks arrive sequentially, task boundaries and identities are known, and each task has a classification head.Source: §2 · p. 3
- The paper considers task-specific subnetworks and memory replay as NCL methods from different approaches.Subnet assigns a pathway to each task, while memory replay replays previous inputs to reinforce old knowledge.Source: §2 · pp. 3–4
- Subnet methods can offer competitive accuracy with relatively small memory and computational overheads, but may be limited by representation capacity.Memory replay often offers high accuracy at the cost of a large replay buffer and extra forward/backward passes.Source: §2 · pp. 3–4
- DRAM is organized hierarchically into channels, ranks, chips, banks, rows, and columns, with banks as the lowest-level units accessible in parallel.A DRAM request selects a rank, accesses chips in that rank, and routes within each chip to a bank and row-column address.Source: §2 · p. 4
3 THE ROBUSTNCL FRAMEWORK
RobustNCL models transient DRAM faults across NCL model states, studies their effects, and uses runtime detect-and-cure to replace anomalous accessed weights with selected safe values.
3.1 NCL TRANSIENT-FAULT MODELING
- The fault model flattens DRAM banks into continuous memory, randomly injects bit-flips at a specified fault rate, and considers faults at different trained-model states.A faulty bit remains until overwritten; faults may affect either the initial offline-trained model or the latest NCL-trained model.Source: §3.1 · pp. 4–5, Figure 5 · p. 5
3.2 NCL FAULT-TOLERANCE ANALYSIS
- Bit-flips can alter layer or subnetwork weight distributions, disrupt neural behavior, and make untrained Subnet weights difficult to train through bad initialization.Clean-model profiles include minimum, maximum, and most-probable weight values.Source: §3.2 · p. 5
3.3 LIGHTWEIGHT DETECT-AND-CURE (D&C) STRATEGY
- The framework monitors each DRAM-accessed weight for anomalous values and replaces detected faults with safe curing values.The curing library records per-layer statistics and updates them after NCL training; its values are stored in radiation-hardened memory.Source: §3.3 · pp. 6–7
- Curing variants replace anomalies with zero, a distribution-bound minimum or maximum, or the most-probable value; Subnet statistics are recorded per subnetwork and layer.The library uses offline-trained statistics initially and updates them as runtime NCL training changes the model.Source: §3.3 · pp. 6–7
4 EXPERIMENTAL SETUP
The evaluation tests RobustNCL in task-incremental learning with Subnet and memory replay across four split-image benchmarks and multiple fault rates.
EXPERIMENTAL SETUP
- Task 1 is learned offline, while subsequent tasks are learned sequentially during runtime NCL training.The framework is implemented in Python and evaluated on a single Nvidia RTX 4090 GPU.Source: §4 · p. 7
NCL METHODS AND BENCHMARKING DATASETS
- Experiments evaluate Subnet and memory replay on Split MNIST, Split Fashion MNIST, Split CIFAR-10, and Split CIFAR-100.The first three benchmarks have five two-class tasks; Split CIFAR-100 has ten ten-class tasks.Source: §4 · p. 7
NETWORK, HARDWARE, AND FAULT RATE SETTINGS
- The study uses a VGG9 network and a single DDR3-1600 2Gb DRAM chip, with fault rates varying by method and dataset.Training settings differ between five-task and ten-task benchmarks; MR also uses a 2,000-sample buffer and replay batches of 48.Source: §4 · p. 7
5 RESULTS AND DISCUSSION
Across the tested NCL methods and datasets, RobustNCL preserves accuracy under high transient-fault rates and incurs low processing, memory, and energy overheads.
5.1 MAINTAINING HIGH ACCURACY OF NCL UNDER TRANSIENT FAULTS
- Baseline accuracy drops by more than -10% at fault rates as low as 5×10^-8 to 1×10^-7 for Subnet and 1×10^-5 to 5×10^-5 for memory replay on five-task datasets.The reported thresholds vary by dataset; the ten-task results also report drops above -10% at fR ≥ 1×10^-7 for Subnet and fR ≥ 1×10^-6 for MR.Source: §5.1 · p. 8
- At high fault rates, the best RobustNCL variant stays within -0.5% accuracy of fault-free models across the evaluated methods and datasets.RobustNCLv2 generally performs slightly better on five-task datasets, while RobustNCLv3 generally performs slightly better on the ten-task dataset.Source: Table 1 · p. 8, §5.1 · p. 8, Table 2 · p. 9
5.2 OVERHEADS OF ROBUSTNCL
- Up to 2.3% latency overhead, less than 1% memory overhead, and 6.35% average energy overhead are reported across methods, datasets, and fault rates.The authors attribute the costs to per-access detect-and-cure processing and updates to the curing library after NCL training.Source: §5.2 · p. 9
6 CONCLUSION
RobustNCL detects anomalous DRAM weights during access and cures them using three replacement variants; the paper reports high accuracy under faults with low overhead.
6 CONCLUSION
- RobustNCL detects anomalous weights during DRAM access and replaces them with zero, min/max, or most-probable values from the clean weight distribution.The conclusion reports accuracy within -0.5% of fault-free models for the best variant, with negligible latency and energy overheads.Source: §6 · p. 9