www.mathworks.com/help/stats/pca.html Take a look at the "percent variability explained" example, it should answer your question with code :) goodluck!

Honestly, you can fill an entire course with answer everything about this properly. I think the most important thing for you to understand is something called "hyperparameter tuning". A standard (and sadly, basically best practice approach) is to write down every variation of each of those, and then systematically try them. Sound primitive? It is. Lots of alternatives are being tried, but I've yet to see one turn out systematically better, despite some papers claiming it does for a specific problem here or there. It's a big part of why ML takes up so much CPU/GPU resources. As an example, I will usually create a network of different depths, and for a CNN make a number of filters per layer, and width of filters per layer, and then run a few hundred variations. Doing it randomly can help find the best possible approximation of the global minimum. I will usually seperatly before and after try things like optimizer and optimizer hyper parameters - sometimes people call this quasi manual approach "grad student decent" because it involves the sort of time only a grad student has. Anyway, "hyperparameter tuning" is what you want to google. I talk about it a bit in a few of my videos, if you want to checkout the other, but I don't have a focused video on it yet.

@@CodingLikeMad thank u very much

@@CodingLikeMad the matrix that I report in Matlab must be numeric. Should I normalize it or do I carry the values as they are? Does the neural network vary?

It is good practice to standardize your input features before using them in a model. This mean making them have 0 mean and standard deviation 1. This is most helpful when your values have vastly different magnitudes, because it keeps gradients well defined.

Both classification and regression NN in matlab use backprop algorithms to my knowledge.

I think an image segmentation video(pixel classification specifically) would super interesting! Thanks for the suggestion. Not sure when I'll get to it, but I want to learn how to do make those models for sure. Satellite imagery datasets sound like a ton of fun to play with, to be honest. I wonder if there is a public set to play with?

Hi Ahmed, if you wanted to make something like that (and I don't know of an existing neural network for this, although there may be one), you need to consider what your input and output features are. Classification networks are good for true/false questions there (like water quality bad vs. good), but would not be helpful for things like predicting heavy metal levels (a continuous number). For something like that I suggest a regression network. I give an example of such a network here: kzfaq.info/get/bejne/opxpf9GQtdDGZZc.html For a problem like this, I think the three things you need to work out are: 1. What input features would you like to use, and how to express them. For instance, you might include an indicator for what the distance to heavy industry or agriculture is, but how do you work that number out? 2. What would you like to predict? How do you define water quality? 3. What data set do you have that is sufficient to train the network? I would actually start with 3, and look at what is available. To be valuable, I guess that it should have features you can generate WITHOUT a measurement, such as the air quality in the region from a weather station, or the amount of water flow, or whatever else would be helpful (I have no idea), rather than just different aspects of the water quality itself. In any case that is a very interesting project, good luck going forward! :)

thank you for replying. For the neural network project that monitors water well or bad. The information input is (PH - Conductivité- Chlorures - Sulfates- Calcium-Magnesium -Fer). The output is 1 for good water. Or 0 for bad water.

Hi Ahmed, a follow up: for this case you need a training set where someone has labeled the data as good or bad. If you have that, then the techniques here can be applied directly, just use 0 and 1 for the color labels(more or less), and expand the input features to be the list you have instead of rgb. Good luck!

Thank you very much for your help .... If you have problems I will communicate with you.

Sorry for the delay in response. With only 200 training examples, my suggestion would be to consider not using a neural net. If you do, the layers must be very limited to avoid overtraining. This is observed by having a high accuracy on your train set, and low accuracy on your test set. I suggest considering a decision tree or random forest instead, with limited number of nodes. A cross validation set is critical to prevent over training.

@@CodingLikeMadplease ; i need to talk you in email

@ezziane hassan I've now added a contact email to my about section if you would like. I'm not sure how much help I can give, but I'll try :)