Overcoming Catastrophic Forgetting: Continual Learning for Radar Waveform Classification

Mohamed Rabie; Chinthana J. Panagamuwa; Konstantinos G. Kyriakopoulos

eess.SPOct 13, 2026

Overcoming Catastrophic Forgetting: Continual Learning for Radar Waveform Classification

Authors: Mohamed Rabie, Chinthana J. Panagamuwa, Konstantinos G. Kyriakopoulos

Organizations: Wolfson School of Mechanical, Electrical and Manufacturing Engineering, Loughborough University, U.K.

Abstract

Automated radar waveform recognition classifies the waveform and modulation type of a received signal. As new waveform variants appear over time, a deployed classifier must learn them without retraining from scratch. Catastrophic Forgetting occurs when learning new waveforms overwrites prior knowledge. Continual Learning (CL) methods tackle this challenge, but prior radar studies only evaluate a few methods on narrow benchmarks. Practitioners lack evidence on which method to deploy or which mechanisms are effective. This work presents the first large-scale benchmark of continual learning for radar waveform recognition. Eighteen CL methods are evaluated on a sequential ten-task problem, drawn from four public datasets spanning 54 waveforms, under task- and class-incremental (TIL, CIL) settings and single- and multi-epoch regimes. Methods that retain data from old tasks lead both TIL and CIL settings. Plain replay alone closes 77-88% of the gap between the lower and upper practical limits under TIL, but only 46-55% under CIL. The internal mechanisms of the leading methods are ablated and statistically tested to identify effective mechanisms. Each ablation is paired against its full method over n=6 training seeds, topped up to n=12 when inconclusive, and p-values are Holm-Bonferroni corrected within each method's family. Ablation results show that several mechanisms confer smaller but statistically significant gains. Combining these mechanisms improves the leading F1 score by 3.4-5.5% and 2.7-4.9% under TIL and CIL, respectively, and the combined models exceed every benchmarked method in both settings for all SNRs. These results give researchers of radar waveform classification systems a mechanism-level basis for design.

Figures & tables

Approach	Mechanism	Key Strength	Primary Limitation
Regularisation (LwF [ 4 ] , EWC [ 5 ] , SI [ 6 ] , RWalk [ 7 ] , UCL [ 34 ] )	Penalises changes to parameters identified as important for prior tasks in the parameter or feature space	No exemplar storage required; potentially lightweight and broadly applicable across architectures	Importance estimation must be reliable and extensible to new tasks
Replay (Res-ER [ 8 ] , Ring-ER [ 10 ] , iCaRL [ 10 ] )	Interleaves stored exemplars with current training data via cross-entropy replay	Simple and effective; effectiveness scales naturally with buffer size	Buffer storage requirement grows with the number of tasks and classes
Optimisation (GEM [ 9 ] , AGEM [ 11 ] , C-/S-/La-MAML [ 12 ] )	Uses stored exemplars or a replay buffer to constrain or shape the gradient update direction via an explicit optimisation procedure	Principled gradient-level forgetting prevention without relying solely on rehearsal frequency	Constraint projection or bilevel updates introduce significant per-step compute overhead
Architectural (PackNet [ 13 ] , HAT [ 14 ] )	Allocates disjoint parameter subsets per task via pruning or learned masks, preventing any cross-task parameter interference	Eliminates forgetting by ensuring that previously learned tasks are unaffected by subsequent tasks	Fixed network capacity imposes a hard limit on the number of disjoint subnetworks.
Hybrid (CTN [ 33 ] , BCL-Dual [ 32 ] , La-ER [ 12 ] )	Combines replay and optimisation-based updates that decouple stability and plasticity signals across separate memory streams	Complementary mechanisms can yield better stability–plasticity trade-offs than either alone	Increased architectural and training complexity relative to single-mechanism approaches

TABLE I: Summary of CL families by mechanism, key strength, and primary limitation.

Signal Type / Family	DeepRad	RadNIST	RadChar	RadarML
Non-modulation	✓	✓		✓
FM (LFM, FMCW, NLFM, HFM, QFM)	✓	✓	✓	✓
Coherent pulse train			✓
FSK (2/4/8-FSK)	✓			✓
PSK (2/4/8-PSK)	✓
Costas code	✓

TABLE II : Signal types available in evaluated datasets.

Dataset	Task	Signal Types (Classes)	# Classes	Size	Difficulty
RC	1	Coherent Pulse Train, Barker Code, Polyphase Barker Code, Frank Code, LFM	5	50k	$97.4\text{\tiny$ \pm 0.1 $}$
RN	1	Q3N#2, Q3N#1, Q3N#3, P0N#2-P3, P0N#2-Frank, P0N#2-Px	6	60k	$90.1\text{\tiny$ \pm 0.4 $}$
RN	2	P0N#1, P0N#2-Barker, P0N#2-P2, P0N#2-P4, P0N#2-P1, P0N#2-Zadoff-Chu	6	60k	$96.0\text{\tiny$ \pm 0.2 $}$
RML	1	FSK #0, Barker-5, Barker-13, Barker-11, FMCW	5	50k	$81.7\text{\tiny$ \pm 0.9 $}$
RML	2	Barker-5 ${\times}$ 11, Barker-5 ${\times}$ 13, Barker-7, FSK #3, CW	5	50k	$84.5\text{\tiny$ \pm 0.4 $}$
RML	3	FSK #1, QFM , HFM , FSK #2, NLFM	5	50k	$73.9\text{\tiny$ \pm 1.2 $}$

TABLE III : Class partitions per task across datasets. Size indicates number of samples before training/testing split. Difficulty is the F1 (%) of a model trained on that task alone for

5

epochs; lower values indicate harder tasks.

Hyperparameter	Value
Architecture	ResNet18-1D
Optimiser	SGD
Epochs	1 (single-epoch) / 5 (multi-epoch)
Batch Size	256
Replay Batch Size	128
Replay Memory Size	5120

TABLE IV : Shared hyperparameters. Replay parameters are only applicable to methods that use a memory buffer.

Panel of Table 5 — TABLE V : Continual learning performance for the TIL setting under 1E and 5E training. Results show the mean and sample standard deviation (in $\%$ ) over training seeds. $\Delta\bar{F1}_{\text{CL}}$ is the change in $\bar{F1}_{\text{CL}}$ from 1E to 5E .

Continual Learning

Continual Learning studies how models acquire knowledge from changing data streams while retaining or selectively revising what they learned before.

See all 301 papers

ELC: Evidential Lifelong Classifier for Uncertainty Aware Radar Pulse Classification
Apr 8, 2026Mohamed Rabie, Chinthana Panagamuwa, Konstantinos G. KyriakopoulosClassification Uncertainty Quantification
Wolfson School of Mechanical, Electrical and Manufacturing Engineering, Loughborough University, U.K.
Same Methods, Different Rankings: Trainable Depth as an Evaluation Variable in Continual Learning
Apr 23, 2026Paul-Tiberiu Iordache, Elena Burceanu, Mihai DascaluTask-Incremental Learning Fine-Tuning
Bitdefender, Romania · Politehnica University of Bucharest, Romania
SPARCL: Spectral Partitioned Analytic Continual Learning
Aug 21, 2026James Hartley, Zeropy Surio, Daniel Whitmore +2Continual Learning Class-Incremental Learning
University of Sheffield

Apr 8, 2026J/K move · Enter open · S save

ELC: Evidential Lifelong Classifier for Uncertainty Aware Radar Pulse Classification

Mohamed Rabie, Chinthana Panagamuwa, Konstantinos G. Kyriakopoulos

Wolfson School of Mechanical, Electrical and Manufacturing Engineering, Loughborough University, U.K.

Classification Uncertainty Quantification Evidential Deep Learning Online Continual Learning Selective Prediction

Open paper AskPDF arXiv

Reliable radar pulse classification is essential in Electromagnetic Warfare for situational awareness and decision support. Deep Neural Networks have shown strong performance in radar pulse and RF emitter recognition; however, on their own they struggle to efficiently learn new pulses and lack mechanisms for expressing predictive confidence. This paper integrates Uncertainty Quantification with Lifelong Learning to address both challenges. The proposed approach is an Evidential Lifelong Classifier (ELC), which models epistemic uncertainty using evidence theory. ELC is evaluated against a Bayesian Lifelong Classifier (BLC), which quantifies uncertainty through Shannon entropy. Both integrate Learn-Prune-Share to enable continual learning of new pulses and uncertainty-based selective prediction to reject unreliable predictions. ELC and BLC are evaluated on 2 synthetic radar and 3 RF fingerprinting datasets. Selective prediction based on evidential uncertainty improves recall by up to 46% at -20 dB SNR on synthetic radar pulse datasets, highlighting its effectiveness at identifying unreliable predictions in low-SNR conditions compared to BLC. These findings demonstrate that evidential uncertainty offers a strong correlation between confidence and correctness, improving the trustworthiness of ELC by allowing it to express ignorance.

Abstract

Continual Learning

Explore similar work

ELC: Evidential Lifelong Classifier for Uncertainty Aware Radar Pulse Classification

Same Methods, Different Rankings: Trainable Depth as an Evaluation Variable in Continual Learning

SPARCL: Spectral Partitioned Analytic Continual Learning