วีดีโอ

Glad it was helpful!

Most welcome!

How can I help you?

Glad it was helpful.

Content is not missing. Its a bit short , That's all.

The formulation has been the premise of why variance reduces theoretically when ensembling is in place compared to independent models. From practical standpoint, it works well which is why random forest is such a star with so many hyperparams to ensure you get different trees as much possible across 100s of features faced in real applications.

You are welcome

Which normalization would you like to see? The wgt computation in each iteration is normalized. Could you clarify.

You are welcome

Glad it was helpful!

You are welcome!

I am glad it was helpful.

I am glad. It was helpful.

I am glad. It was helpful.

Glad you found it helpful.

Yes, you can do that and make it easier to use in multiple places.

Glad it was helpful!

In general, I recommend that we set epochs to very large value say 50,000. Then in your code you setup early exit logic as part of training. This will work best for most cases since the training fit will automatically exit when convergence has happened. Hope that helps.

Thanks for sharing! I am glad the videos are helpful :)

Glad you liked it!

Lets see if can correlate it with a hypotheses that humans would do to learn. Lets say we are in a Forest & searching for trails of human foot marks to get out of it. Every time we find a footprint, we valid & learn about surroundings, vegetation, terrain,etc. Over a period of time we learn ehat leads to exit And what doen't. That precisely the idea here. Hope that helps.

@@machinelearningmastery I'm sorry but I still don't get it. You can explain it with more math. What I don't get is after predicting a miu, why do we need to add omega? Like what does omega do where?

You're welcome

Thank you

Great Question. A. First, there are two weights in the system - one driving the weight of data point and another driving weight of the classifier. I shall explain both in a second. B. Second, the fact is error >0.5 will give negative lambda. This is part of the design as I shall clarify below. So, when error<0.5 then lambda >0; when error = 0.5, then lambda =0; when error > 0.5 then lambda <0. This means, the model actually goes nowhere when error is exactly 0.5. This is why in implementations, they creatively make sure it won't match this magic number and break the model. Now, back to the first point about weights. Let me give with examples. Say error is 0.1, then Lambda=1.10. Wgt for correctly classified data point is 0.33 weight for misclassified data point=3.0; Weight of this classifier is 1.10 (the lambda itself) Say error is 0.9, then Lambda=-1.10. Wgt for correctly classified data point is 3.0 weight for misclassified data point=0.33; Weight of this classifier is -1.10 (the lambda itself) If we see above, few things are clear 1. If a classifier is classifying well (with low error rate), its overall weight in the ensemble stack in positive. And whatever misclassification is there is given priority in the next run. 2. If a classifier is poor (with high error rate), its overall weight in the ensemble stack is deeply negative. And whatever classification happened correctly, we continue to keep that going for the next run with the hope we can improve its influence gradually. Hope that clarifies. Let me know if there is still an open point on this.

Thank you.

Glad to know it was helpful. :-) please spread the word.

Typo. Thanks for highlighting. Shall update notes

Yes, bias term do exist. Remember whenever you see W matrix in the optimization space, W = [w,b]. The optimizer is working on W as a whole, so its a parallel update so that it can calculate the actual loss and minimize it. I will try to see if I can do a video to illustrate this point. Also, as you progress in your ML study, you may come across models like RNN etc. -- in such cases the W = [w,u,b]. Therefore, to generalize our learning, it is important to see that W is a matrix that contains all the weight-units (includn bias) that the optimizer must search and input into the system to get a loss. Hope that clarifies.

@@machinelearningmastery I see, what I meant was the video we have function = wx but if it was possible to do function = wx + b, I think it is the same but just another term added

Yes, that is correct , bias would be another parameter to be found by GD along with all other weights it finds for each input features.

For use in say python snippet or excel to get similar results as in my video, you may use a small value like 0.01 to start off. Such small values should not impact your convergence time. If we use large values (and say we got it wrong), it would take longer to converge. But, what happens in practice with NeuralNet training is -- use the first 10 or 25 epochs as "warm up" epochs. Generally, large learning rate is used during the warmup epochs and purpose of warm up is to get some of the parameters initialized. Following that, actual training would start. Hope that clarifies.

If you are on Keras, its a 2 liner, Let me give you code here: #fit the model and capture return value of history history = model.fit(X_train, y_train, epochs=epochs,.....) #plot how validation vs training curves look using the "history" variable. plt.plot(history.history['accuracy'], label='train') plt.plot(history.history['val_accuracy'], label='test') plt.legend() plt.show()

I have gone over the code in the video. Any particular segment not clear?