Hello everyone,
I have applied 1D CNN on speech emotion recognition, when I shuffled columns I got diffrent results, for example, using matrix(:,[1 2 3]) gives different classification results than matrix(:,[2 3 1]) which should be the same, I have tried rng("default") but it hasn't resolved the issue. Can someone please assist me with this?
Thank you in advance!
Hamza Roubhi
This is an interesting observation. Getting different classification results when shuffling columns of the feature matrix implies a couple issues:
1. The model is likely overfitting on spurious order-dependent patterns rather than generalizable features. This can happen with insufficient training data.
2. Shuffling may be disrupting temporal relationships between features that the 1D CNN relies on. Emotion has sequential structure.
Some ways to address this:
- Increase the training set size and diversity to reduce overfitting.
- Introduce regularization techniques like dropout or weight penalties to improve generalizability.
- Try a 2D CNN or recurrent neural network architecture better suited to spatiotemporal data instead of 1D CNN.
- Ensure sufficient shuffling during training itself - the model should learn invariance to column order.
- Analyze if certain columns are disproportionately impacting performance when shuffled.
The goal should be improving model robustness so it maintains performance regardless of input feature order.
Hamza Roubhi I appreciate your dedication to applying 1D CNN in the domain of speech emotion recognition and your commitment to addressing the issue concerning the variation in classification results when shuffling columns. As a fellow researcher with a background in signal processing, I understand the significance of consistency and reliability in the outcomes of such analyses.
When encountering discrepancies in results due to column shuffling, it is essential to ensure that the underlying data preprocessing and feature extraction methods remain consistent across different column arrangements. Validating the integrity of the data and confirming that the shuffling process does not introduce unintended variations can help maintain the robustness and reliability of the classification results.
Additionally, verifying the implementation of the CNN architecture, including the configuration of the layers, activation functions, and training parameters, is crucial in ensuring reproducibility and consistency in the classification outcomes. Reviewing the model's initialization procedures and ensuring that the randomization process aligns with the desired standards of consistency can potentially address the issue you are facing.
Furthermore, conducting thorough checks on the data partitioning and validation procedures, such as cross-validation techniques, can help identify any potential sources of variability that may arise due to the shuffled column arrangements. Ensuring that the model training and testing processes maintain consistency and adhere to standardized protocols can contribute to the stability and reliability of the classification results.
I recommend conducting a systematic evaluation of the data preprocessing, model architecture, and training protocols to pinpoint any potential sources of variation that may arise during the column shuffling process. By adhering to rigorous validation practices and ensuring the reproducibility of results across different column permutations, you can enhance the reliability and robustness of your CNN-based classification framework for speech emotion recognition.
- Badges
- Science topic