FuseAMC
A lightweight decision fusion model for automatic modulation classification that combines a residual CNN with a gradient-boosted tree classifier based on higher-order cumulant features. A logistic regression meta-combiner fuses the posterior probabilities of the two, and the model can be trained and run for inference on a single CPU core.
Timeline (last 90 days)
FuseAMC is a lightweight decision fusion model that achieves performance surpassing each of its baselines on the RadioML 2016.10a dataset, and the entire pipeline requires only a single CPU core to train and run.
The core of FuseAMC is a logistic regression meta-combiner that takes the concatenated class posterior probabilities of the two sub-models as input and is trained on a validation set disjoint from the test set.
FuseAMC's two sub-models are a residual CNN with about 198k parameters and a gradient boosting tree (GBT) cumulant classifier.
The entire FuseAMC model has about 0.2 million parameters, a model footprint of only a few megabytes, and about 3 MFLOPs of computation per frame.
On the RadioML 2016.10a dataset, using 5 random seeds and a held-out test set of 44k frames, FuseAMC achieves 55.8% overall accuracy.
FuseAMC improves by 5.1 percentage points over the best cumulant-feature baseline and by 3.2 percentage points over the standalone CNN.
Under the paired-bootstrap 95% confidence interval, the lower bounds are 5.0 and 3.1 percentage points respectively, both positive, indicating the improvements are statistically significant.
FuseAMC matches the performance of the best classical baseline in the high-SNR range and matches the CNN's performance in the low-SNR range.
All Facts (8)
FuseAMC is a lightweight decision fusion model that achieves performance surpassing each of its baselines on the RadioML 2016.10a dataset, and the entire pipeline requires only a single CPU core to train and run.
50%UnverifiedThe core of FuseAMC is a logistic regression meta-combiner that takes the concatenated class posterior probabilities of the two sub-models as input and is trained on a validation set disjoint from the test set.
50%UnverifiedFuseAMC's two sub-models are a residual CNN with about 198k parameters and a gradient boosting tree (GBT) cumulant classifier.
50%UnverifiedThe entire FuseAMC model has about 0.2 million parameters, a model footprint of only a few megabytes, and about 3 MFLOPs of computation per frame.
50%UnverifiedOn the RadioML 2016.10a dataset, using 5 random seeds and a held-out test set of 44k frames, FuseAMC achieves 55.8% overall accuracy.
50%UnverifiedFuseAMC improves by 5.1 percentage points over the best cumulant-feature baseline and by 3.2 percentage points over the standalone CNN.
50%UnverifiedUnder the paired-bootstrap 95% confidence interval, the lower bounds are 5.0 and 3.1 percentage points respectively, both positive, indicating the improvements are statistically significant.
50%UnverifiedFuseAMC matches the performance of the best classical baseline in the high-SNR range and matches the CNN's performance in the low-SNR range.
50%