PulseExploreJournal ClubDebatesTrendingResearchersJournals
Instagram
HomeExploreJournal ClubTrending
Synapse
⌘+K
Synapse
March 16, 20260 citationsOpen Access

Training autoencoders on their own outputs causes collapse

TSThomas Schanze

Key Result

Training autoencoders on their own outputs caused a drastic drop in performance, leading to nearly complete correlation of outputs after approximately 100,000 steps.

Key Points

  • This study aims to investigate the effects of augmenting training data with autoencoder-generated outputs on model performance.
  • Trained classical autoencoders and denoising autoencoders on ECG signal data.
  • Generated reconstructions of the input data for potential data augmentation.
  • Evaluated performance by comparing results with and without augmented data.
  • Augmenting the training set with autoencoder outputs led to a significant drop in model performance.
  • The performance collapse was categorized as catastrophic, highlighting risks in using generated data.

Structured PICO

P
Population
300 ECG segments from 30 healthy patients (10 segments per patient), each containing P-wave, QRS-complex, and T-wave (95 datapoints per segment).
I
Intervention
Recursive data manipulation where autoencoder (AE) or denoising autoencoder (DAE) generated reconstructions are added back into the training set.
C
Comparator
Autoencoders trained only on original reference data without data feedback.
O
Outcome
Average pairwise correlation index between network model outputs to measure network collapse.

Training autoencoders on their own generated ECG data causes catastrophic performance collapse, demonstrating the risks of AI data self-contamination.

Limitations

  • Self-contamination of training data potentially limits generalizability.
  • Only 30 healthy patients were used for the ECG segments.
  • The approach used for augmenting data is an exaggerated variant of possible contamination.

Abstract

Classical autoencoders (AE) learn a compressed, meaningful representation of the input data and denoising autoencoders (DAE) capture the true underlying data manifold even when inputs are noisy. Data is the foundation of artificial intelligence, and thus for all autoencoder types. However, all types produce, when well trained, output data which are similar to the input data. This could lead to output data being added to the data that is to be used for further learning. We show on ECG signals that adding AE/DAE-generated reconstructions to the training set — intended to augment data — causes catastrophic performance collapse.

Ask AI
Helpful
Bookmark
Share
View Full Paper

Cite This Study

Thomas Schanze (2025) studied ECG signal processing (n=300). Denoising Autoencoder (DAE) vs. Classical Autoencoder (AE) was evaluated on Average pairwise correlation index between network model outputs. Training autoencoders on their own outputs caused a drastic drop in performance, leading to nearly complete correlation of outputs after approximately 100,000 steps.

synapsesocial.com/papers/69b79e538166e15b153ab787https://doi.org/10.18416/automed.2026.2470
Ask AI
Helpful
Bookmark
Share
View Full Paper

Also Consider

Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context:

  1. 1Rethinking Autoencoders for Medical Anomaly Detection from A Theoretical Perspective2024 · 1 citations
  2. 2A comprehensive study of auto-encoders for anomaly detection: Efficiency and trade-offs2024 · 46 citations
  3. 3ForTIFAI: fending off recursive training induced failure for AI model collapse2026
  4. 4Original Research Article Enhancing Missing Data Imputation with Improved DAE Training and Input Recombination2024 · 1 citations
  5. 5Extracting and composing robust features with denoising autoencoders2008 · 7,423 citations