Dear Afolabi Ifedayo Awodeyi

You can use matlab to design and optimize your convolutional neural networks.

You can run sufficient experiments to decide for the network parameters which gives you sufficient classification or identification accuracy.

You can make a compromise between accuracy and complexity. You can make your optimization experimentally.

Please read the following article it will give you a understanding of different optimizers:

1. https://towardsdatascience.com/optimizers-for-training-neural-network-59450d71caf6

2. Conference Paper Performance comparison of the convolutional neural network o...

Dear Afolabi Ifedayo Awodeyi thank you for this question

The links of Gyanendra Chaubey are really helpful. How about using Phyton in this case?

Dear Leila Nasraoui , and Afolabi Ifedayo Awodeyi

In case of Python, you can use the hyperparameter tuning techniques like GridSearchCV, RandomSearchCV, Genetic Algorithms, Hyperopt, and Optuna etc. to decide the hyperparameters (1). These techniques will also help you in the optimization of your model. For the different Optimizers used for the neural networks you can check the variation of results manually using tensorflow (2) after optimizing the parameters.

1. https://scikit-learn.org/stable/modules/grid_search.html#randomized-parameter-optimization

2. https://www.tensorflow.org/api_docs/python/tf/keras/optimizers

Thank you.

Hope it will work for you.

Dear Afolabi Ifedayo Awodeyi ,

This question could be addressed from different angles. First of all, I strongly recommend using fastai (PyTorch), as it already uses the state-of-the-art optimization techniques. Second, you could start by training on a huge face recognition dataset without masks, and then fine-tune on you small subset with masks. This will improve a lot the performance of the model.

It would also help to understand which kind of problem you are solving. I guess it is an object detection, and therefore, a model like YOLO would perfectly work for your task. You can find a nice implementation here: https://github.com/ultralytics/yolov5

It is usually not predictable which hyperparameters will work best with your model on your specific data. These are typically chosen based on the model performance on a validation set.

I have applied some optimisation techniques for deep learning hyper-parameters tuning. I guess, it might be useful to see

1. Article A deep learning-based evolutionary model for short-term wind...

Preprint Hybrid Neuro-Evolutionary Method for Predicting Wind Turbine...

Tensorflow (with keras) or pytorch calculate the gradient descent for you. Also provide several types of layers for your CNN.

You may also follow https://towardsdatascience.com/4-pre-trained-cnn-models-to-use-for-computer-vision-with-transfer-learning-885cb1b2dfc and https://machinelearningmastery.com/how-to-use-transfer-learning-when-developing-convolutional-neural-network-models/