PaperlayerView exact revision on arXiv

Source-linked AI summary

SR2CNN: Zero-Shot Learning for Signal Recognition

Yihong Dong, Xiaohan Jiang, Huaji Zhou, Yun Lin, Qingjiang Shi

arXiv:2004.04892v7cs.LG

TL;DR

Signal zero-shot recognition addresses the problem of classifying signal categories that lack training data, a setting made difficult by signals’ obscure semantics. SR2CNN learns semantic signal features using classification, center, and reconstruction objectives, then uses those features to discriminate classes. The authors report recognition of known and unknown classes and gradual improvement through updates to semantic center vectors.

  • Problem

    Signal ZSL has been little studied because signal semantics are obscure and high-level semantic attributes are difficult to obtain, even though unknown classes may need recognition.

    Source: §I · p. 1

  • Method

    SR2CNN learns semantic signal features with cross-entropy, center, and reconstruction losses, then uses a distance metric to discriminate classes.

    Source: §I · p. 2, A. Feature Extractor, Classifier and… · p. 5

  • Results

    Experiments report that SR2CNN discriminates known and unknown classes and can gradually improve itself.

    Source: §I · p. 2, §VII · p. 12

  • Takeaways & Limitations

    SR2CNN is reported to work when some classes lack sufficient training data, and the paper provides SIGNAL-202002 with eight digital and three analog modulation classes for further research.

    Source: §I · p. 2, §VII · p. 12

Abstract

from arXiv · show

Signal recognition is one of significant and challenging tasks in the signal processing and communications field. It is often a common situation that there's no training data accessible for some signal classes to perform a recognition task. Hence, as widely-used in image processing field, zero-shot learning (ZSL) is also very important for signal recognition. Unfortunately, ZSL regarding this field has hardly been studied due to inexplicable signal semantics. This paper proposes a ZSL framework, signal recognition and reconstruction convolutional neural networks (SR2CNN), to address relevant problems in this situation. The key idea behind SR2CNN is to learn the representation of signal semantic feature space by introducing a proper combination of cross entropy loss, center loss and autoencoder loss, as well as adopting a suitable distance metric space such that semantic features have greater minimal inter-class distance than maximal intra-class distance. The proposed SR2CNN can discriminate signals even if no training data is available for some signal class. Moreover, SR2CNN can gradually improve itself in the aid of signal detection, because of constantly refined class center vectors in semantic feature space. These merits are all verified by extensive experiments.

I. INTRODUCTION

Signal zero-shot learning is challenging because signal semantics are difficult to describe, despite the need to recognize classes absent from training. SR2CNN learns signal semantic features for recognition, and the paper situates this approach against prior closed-set, open-set, and image ZSL work.

I. INTRODUCTION

  • Signal ZSL must recognize unknown classes, but unlike image tasks, signal semantics resist high-level description and semantic attributes are difficult to acquire.The paper distinguishes its multi-class unknown-signal recognition goal from the vibration fault-diagnosis ZSL definition in, which it characterizes as domain adaptation.Source: §I · p. 1
  • SR2CNN combines a CNN feature extractor with an autoencoder to learn semantic signal features, then uses a distance metric to discriminate classes.Its overview also describes maintaining decoder reconstruction and classifier prediction while learning the semantic representation.Source: Figure 1 · p. 1, §I · p. 2, Figure 2 · p. 2
  • Experiments report that SR2CNN discriminates known and unknown classes and can gradually improve itself.The conclusion attributes improvement to updating semantic center vectors.Source: §I · p. 2
  • SIGNAL-202002 is introduced as a new dataset containing eight digital and three analog modulation classes.The paper offers it for further research.Source: §I · p. 2

II. RELATED WORK

  • Prior signal-recognition work largely follows the closed-set assumption, while open-set methods can detect unknown samples but cannot classify among unknown signal classes.The paper says ZSL can classify unknown classes in image tasks, where semantic attributes can be described, but signal attributes are not readily acquired.Source: §II · p. 3

III. PROBLEM DEFINITION

The problem is to learn from labeled known-class signals and predict labels across both known and unknown classes. SR2CNN builds a semantic representation and then applies a distance metric to discriminate among classes.

III. PROBLEM DEFINITION

  • Training data contain labeled signals from known classes, but the target is to predict labels from both known and unknown classes.Unlike open-set learning that only separates known from unknown, this formulation predicts a label in the full class set.Source: §III · p. 3
  • The framework maps signal inputs into semantic space and maps that encoding to a class label, using functions F and P.The stated task is to choose F and P so the framework identifies known and unknown signal classes.Source: §III · pp. 3–4

IV. PROPOSED APPROACH

  • SR2CNN first builds signal semantic space with feature extractor F, classifier C, and decoder D, then uses a distance metric and discriminator P to classify known and unknown classes.F projects signals into latent features; C predicts labels and D reconstructs signals from those features.Source: §IV · p. 4

A. Feature Extractor, Classifier and Decoder

SR2CNN builds a signal semantic space with a feature extractor, classifier, and decoder, training them with classification, center, and reconstruction losses. The decoder’s reconstruction objective helps retain signal information in the semantic features.

B. Discriminator

The discriminator first separates known from unknown signals using distances to known-class semantic centers, then assigns unknown signals to existing or newly recorded unknown classes. It refines unknown-class centers as further signals are assigned, allowing SR2CNN to distinguish multiple unknown classes rather than grouping them all as one.

  • Class-center distance can use Mahalanobis, Euclidean, or diagonal and scalar covariance-based metrics; the latter covariance-dependent choices require class covariances in advance.Source: B. Discriminator · p. 5
  • The discriminator first compares a test signal’s minimum distance to known-class centers with Θ1, using a Gaussian feature assumption and a generalized three-sigma threshold.Source: B. Discriminator · pp. 5–6
  • For a signal judged unknown, SR2CNN creates a new label if none is recorded; otherwise, it compares d2 with Θ2 to decide whether the signal belongs to an existing or new unknown class.Θ2 is selected between d1 and the median distance dmed in the reported numerical testing.Source: B. Discriminator · p. 6
  • Each assigned unknown signal is saved with its group, and the corresponding semantic center is updated as predictions accumulate.Source: B. Discriminator · p. 6
  • SR2CNN distinguishes n known classes and nR recognized unknown classes, whereas common open-set methods group all unknowns into one additional class.Source: B. Discriminator · pp. 6–7

V. EXPERIMENTS AND RESULTS

The experiments evaluate SR2CNN on three signal datasets using selected unknown classes and separate known and unknown evaluation sets. The setup compares zero-shot learning with supervised learning and uses validation-based hyperparameter selection.

V. EXPERIMENTS AND RESULTS

  • Experiments cover datasets 2016.10A, 2016.10B, and 2016.04C.The paper identifies 2016.04C as a counterpart to the other two datasets.Source: §V · p. 7
  • The reported setup filters out samples with SNR below 16, leaving one-tenth of the original data.Hyperparameters are selected by grid search on the validation set.Source: §V · pp. 7–8
  • Table II contrasts supervised and zero-shot results across three datasets, marking the known/unknown boundary and identifying isotopic-class accuracy pairs.Italicized values support transverse comparisons; dashes indicate no corresponding result.Source: Table II · p. 8
  • Unknown classes are selected from classes whose full-supervision accuracy exceeds 50%.Candidates are AM-SSB and GFSK for 2016.10A and 2016.04C, and CPFSK and GFSK for 2016.10B.Source: §V · p. 8
  • Known-class data are split 70%/15%/15% across training, validation, and test, while unknown classes use a 15% test set.The resulting 2016.10A copy contains 12,600 training, 2,700 known-validation, 2,700 known-test, and 600 unknown-test samples.Source: §V · p. 8

A. In-training Views

During training, known-test softmax accuracy converges to dataset-dependent levels, and distance-based cluster prediction incurs almost no perceptible accuracy loss. The cross-entropy loss also oscillates sharply as center and reconstruction losses intermittently shift the model’s learning focus.

A. In-training Views

  • Known-test softmax accuracy converges to roughly 80% on 2016.10A and 2016.10B and 94% on 2016.04C.The distance-based clustering prediction has almost no perceptible accuracy loss compared with softmax.Source: A. In-training Views · p. 8
  • Cross-entropy loss oscillates sharply during training as center and reconstruction losses intermittently shift SR2CNN’s learning focus.Source: A. In-training Views · p. 8

B. Critical Results

The critical results assess known/unknown discrimination using weighted true rates and dataset-specific discrimination intervals. Performance varies by dataset, with 2016.10B showing a narrower interval and the greatest reported zero-shot accuracy loss.

B. Critical Results

  • The selected zero-shot settings aim to maximize 0.4×TKR+0.6×TUR while keeping known precision above 0.95×UB and isotopic classes at two or fewer per unknown class.The search fixes λ2 at 1 and varies λ1 from 0.05 to 1.0.Source: B. Critical Results · p. 9
  • 2016.10B has the greatest zero-shot loss, exceeding 10% in average total and known accuracy relative to supervised learning.The passage also reports inferior true rates for this dataset.Source: B. Critical Results · p. 9
  • For 2016.10B, the two unknown classes are separately identified into two isotopic classes.Source: B. Critical Results · p. 10
  • A discrimination interval is the range where weighted true rate stays above 80%, and its width quantifies discrimination ability.Source: B. Critical Results · p. 10
  • The weighted true-rate curve maxima are about 0.4, 0.2, and 0.4 across the three datasets, while 2016.10B’s discrimination interval is about one-third as wide as the others.The paper relates the narrower interval to more difficult-to-learn features and relatively poor performance.Source: B. Critical Results · p. 10

C. Ablation Study

The ablation study finds that all three loss functions contribute to SR2CNN, but their effects differ across known and unknown classes. Center loss is especially consequential for unknown-class F1, while L1 loss trades a small known-class gain for a larger unknown-class decline.

C. Ablation Study

  • Removing cross-entropy or reconstruction loss lowers training known accuracy; each contributes about 3% to 5%, while center loss has a weaker effect.Source: C. Ablation Study · p. 10
  • Removing center loss decreases unknown-class F1 by 16.7%, compared with a 1.7% decrease for known-class F1.Removing cross-entropy and reconstruction losses decreases known-class F1 by 1.8% and 2.0%, respectively.Source: C. Ablation Study · p. 10
  • L1 loss raises known-class F1 by 0.8% but lowers unknown-class F1 by 6.4%.The authors conclude that the three loss functions are useful, though they promote SR2CNN differently and in different fields.Source: C. Ablation Study · p. 10

D. Other Extensions

The comparison evaluates SR2CNN against outlier detectors and open-set recognition methods on 2016.10A. SR2CNN has the highest reported true unknown rate and AUC, while performance varies with the unknown classes considered.

D. Other Extensions

  • Known and unknown precision are reported as insensitive to changes in unknown classes, but CPFSK is consistently the hardest unknown class.Performance loss in both TKR and TUR is especially striking when CPFSK and GFSK are both unknown; the authors speculate that similarities to known classes may mislead discrimination.Source: D. Other Extensions · pp. 10–11
  • SR2CNN significantly outperforms the compared outlier detectors and open-set recognition methods in true unknown rate on 2016.10A.Table V presents the comparison; the passage does not provide its cell values.Source: D. Other Extensions · p. 11, Table V · p. 11
  • For unknown class j, the reported unknown rate is the fraction of its samples discriminated as unknown, rather than its accuracy.The experiments consider discrimination tasks, not unknown-class identification tasks.Source: D. Other Extensions · p. 11
  • SR2CNN has the largest AUC among the compared methods, while several methods show a steep false-known-rate cliff near 0.5.The authors speculate that the difficult samples may all be GFSK.Source: D. Other Extensions · p. 11

VI. DATASET SIGNAL-202002

SIGNAL-202002 is a newly synthesized dataset of 11 modulation types, evaluated with CPFSK and B-FM as unknown classes. In that setting, TKR and TUR only slightly exceed 80%, while known-class accuracy remains encouraging relative to supervised learning.

VI. DATASET SIGNAL-202002

  • SIGNAL-202002 contains 11 modulation types, with 20,000 frames per type and 128 samples per frame.The dataset covers eight digital and three analog modulation classes.Source: §VI · p. 11
  • The dataset's synthetic channel model combines additive white Gaussian noise, Rayleigh fading, multipath, and clock offset.Each frame is passed independently through the channel model to emulate real-world signal conditions.Source: §VI · p. 12
  • On SIGNAL-202002, TKR and TUR both just pass 80%, while per-class accuracy remains encouraging compared with the fully supervised model.The selected unknown classes are CPFSK and B-FM.Source: §VI · p. 12
  • The authors speculate that the absence of two unknown classes may unintentionally let SR2CNN focus better on known-class features, leading to superior performance on the known-class classification task.This is a proposed explanation, not an established causal result.Source: §VI · p. 12

VII. CONCLUSION

The conclusion presents SR2CNN as a zero-shot framework for discriminating known and unknown signal classes, including when training data are insufficient for some classes. It also reports gradual improvement through semantic-center updates and introduces a new signal dataset.

VII. CONCLUSION

  • SR2CNN extracts semantic signal features and discriminates both known and unknown classes, including when training data are insufficient for some classes.The authors report that extensive experiments demonstrate its effectiveness.Source: §VII · p. 12
  • The framework can generally improve itself by updating semantic center vectors.Source: §VII · p. 12
  • The paper provides SIGNAL-202002, with eight digital and three analog modulation classes, and identifies complex neural networks as a possible future direction for I/Q signals.Source: §VII · p. 12

APPENDIX A CONVOLUTION AND DECONVOLUTION OPERATION

The appendix expresses convolution and deconvolution as matrix operations and describes their forward and back-propagation forms. It also illustrates convolution with a small input-output example.

APPENDIX A CONVOLUTION AND DECONVOLUTION OPERATION

  • Convolution is represented as multiplication by a sparse matrix M, with the input and output matrices vectorized as a and b.The appendix also gives the corresponding convolution back-propagation expression.Source: Appendix A · p. 12
  • Deconvolution uses a learned matrix with the same shape as M^T in forward propagation, and its transpose in back propagation.Source: Appendix A · pp. 12–13
  • For a 4×4 input and 2×2 output, the vectorized input has 16 dimensions and the output has 4.The example describes M as a 4×16 matrix.Source: Appendix A · pp. 12–13

Related papers

Loading 2004.04892v7…