Hello Saad,

I took a brief look onto the attached paper. As far as I understood, the training/evaluation etc is just the standard procedure for CNNs. The paper uses an Alexnet-like architecture (similar to the literature cited in the introduction), there is nothing special about the architecture, training and evaluation.

The only novelty is the data augmentation scheme the authors are using. The authors treat the speech spectrograms as images and apply a lens distortion to the images, which they call "DAARIP algorithm". Multiple distorted versions (steps 3, 4 and 5) are added to the augmented dataset.

However, I would be careful with the claim to achieve 99% accuracy with this approach. The trick of that paper is that it uses data augmentation BEFORE dividing the dataset into training, validation and test. This means, that an augmented version of a spectrogram can be in the test dataset, while the original is used in the training dataset. So I am not sure whether this paper is only a strong overfitting of the network and inability of the authors to notice this.

You can use this technique to reproduce the result of that paper, however I expect that the performance of that approach will reduce significantly when testing data which is truly unseen to the network.

Regards,

Michael

Luay you didn't read the question! The question is not about a deep learning neural network for detecting the fovea and optic disc! All you did was copy the abstract from "Multiscale sequential convolutional neural networks for simultaneous detection of fovea and optic disc" which is not even pertinent.

Hello Saad,

I took a brief look onto the attached paper. As far as I understood, the training/evaluation etc is just the standard procedure for CNNs. The paper uses an Alexnet-like architecture (similar to the literature cited in the introduction), there is nothing special about the architecture, training and evaluation.

The only novelty is the data augmentation scheme the authors are using. The authors treat the speech spectrograms as images and apply a lens distortion to the images, which they call "DAARIP algorithm". Multiple distorted versions (steps 3, 4 and 5) are added to the augmented dataset.

However, I would be careful with the claim to achieve 99% accuracy with this approach. The trick of that paper is that it uses data augmentation BEFORE dividing the dataset into training, validation and test. This means, that an augmented version of a spectrogram can be in the test dataset, while the original is used in the training dataset. So I am not sure whether this paper is only a strong overfitting of the network and inability of the authors to notice this.

You can use this technique to reproduce the result of that paper, however I expect that the performance of that approach will reduce significantly when testing data which is truly unseen to the network.

Regards,

Michael

Hi Saad,

As Michael, I read also carefully the attached paper. In my opinion, they applied the following process as data augmentation:

1. Applied short-Time Fourier Transform on data to produce spectrogram (this process offers the possibility to transform 1D speech data into 2D data).

2. Applied a lens distortion to the spectrogram. But this step is very poorly explained! There are 3 different type of distortion (bigger, smaller and original picture sizes).

3. All distorted spectrogram are resized at the same size (256 x 256 pixels)

4. This distorted data is randomly divided into 3 datasets (training, validation, testing dataset) as usual.

5. A model with an Alexnet-like architecture is trained on the training dataset.

I share the point of view of Michael on the likely overfitting! If you implement this paper, I am really interested by the results.

Nicolas