huckiyang
commented
3 years ago Dear Yixiong @cyx617,
Thank you for this question. At the beginning of this project, I did not tune the feature but instead increase the data scale. (e.g., set all the data by using --port 1) I only did decode once and used around a month to decode all the acoustic feature in the Google SpeechCommands. But I did observe if the data scale could not represent the full speech dataset, the random sample features would not be enough to train a speech model during the small-scale data re-runs.
For evaluating the stability of acoustic modeling, please use --port 1. But if your task belongs to an image classification, the data scale could be smaller. See https://pennylane.ai/qml/demos/tutorial_quanvolution.html with only 100 examples with random circuit to train the model.
In short, this issue can come from two perspectives from my empirical experience.
-
For the small-scale image data (e.g., MNIST), the issue of model representation is not scale-sensitive but for acoustic model it is better to use all the features to train the model.
-
For the power of quantum circuit and data sampling, there is a still open research problem. Some reference below maybe help, I am also open to collaborate in this kind of problems.
-
Sampling random quantum circuits: a pedestrian’s guide
-
Random Circuit Sampling: Thoughts and Open Problems, Scott Aaronson
Note, we will give QuantumNN Tutorial in IJCAI 2021. Feel free to check.
cyx617
Dear authors.
In a hybrid quantum-classical neural network, the quantum layer composed of parameterized quantum circuit usually integrates with other classical layers (e.g. convolutional layers, dense layers) and the parameters of these quantum circuits are updated together with those of classical layers through classical optimization method like SGD. But in this project, it seems that the parameterized quantum circuits do not directly integrate with other classical layers and they are not trainable. Each time when performing quantum feature extraction from Mel-spectrogram inputs, the parameters of the quantum circuits are randomly initialized and then kept fixed. So the quantum features extracted in this way are also random each time and will not be optimized any more before feeding into the latter parts of the model. Actually, I've run your code (
python main_qsr.py --sr 16000 --port 100 --mel 1 --quanv 1 --eps 100) for a few times and got different model performance each time in terms of converged val accuracy (e.g. 0.19, 0.57). I guess this unstable performance is due to the random quantum feature extracted by quantum circuits. My question is how we could make sure these random features are optimal for the current task without being further optimized. Many thanks!