Based on the code you provided, it seems that you are correctly implementing the hyperparameter tuning using Bayesian optimization and training your network using ‘trainNetwork’ in MATLAB.
However, there can be few potential reasons of why you are not getting test accuracy above 75%:
- Limited model capacity: The network architecture you defined has a shallow structure with a maximum of four fully connected layers. For complex datasets like MNIST, deeper architectures may be necessary to capture more intricate patterns and achieve higher accuracy.
- Insufficient training data: It is important to have enough training data to train a deep learning model effectively. In your code, you are using a 70:30 train-test split, which might result in limited data for training. Consider using more training samples, if possible, to provide the model with a larger and more diverse set of examples.
- Hyperparameter values: Although Bayesian optimization helps in finding good hyperparameter values, it is still possible that the optimal values are not found within the search space defined by your hyperparameters. Experiment with different ranges and values for the hyperparameters to see if you can achieve better results.
It is important to acknowledge that achieving higher accuracy on the MNIST dataset is indeed possible, with many models able to achieve test accuracy well above 90%. So, exploring more advanced network architectures, such as convolutional neural networks (CNNs), can be advantageous, especially for image classification tasks like MNIST.
