Thank you for the vid 👍 But what do you mean by "thresholds" at 11:10 ? Like, what are the thresholds in terms of neural networks, and how can we change them? Thank you :)
Thank you for the positive feedback :) The most simple way to think of a threshold is through a simple model, with only 1 feature as an input and two possible classes as output (say 0 and 1). Then when the model is trained it finds the "best" threshold for the input feature. So, for example, if we denote the input feature as a, then the model may learn that if a>=0.5 then the label is 0, otherwise it's 1. In this example the threshold is 0.5. Neural networks work differently (in most cases) and thus thinking of a threshold may be confusing. For our example from the paragraph above, the output of a neural network will be a vector/list of size 2, where each index is the probability that the output is in that specific class. For example if the output is [0.14, 0.85] then the model "thinks" that there is a 14% chance that the input is from label 0 and a 85% chance its from label 1. If our neural network had only 1 neuron then the 0.5 value from the example above could be incorporated into it. "How do we change it?" - This really depends on what you want to achieve. If FA are more important than FN, or the other way around then you can change your loss function and the incorporated threshold will change accordingly. Hope this helps :)
@@ml_dl_explained cool, thanks 👍 Tbh, I didn't really get how can we "vary the thresholds" to further plot ROC or AUC for a neural network. . I mean, when a model is trained - we have only a single point at the ROC plot (current state of the model). But then how can we "change the thresholds" to have multiple points on the plot? Thank you 😊
You're welcome :) When we train the model, it outputs probabilities. We can change the threshold of those probabilities to get different labels - for example, if the model's output is [0.34, 0.66], one threshold could be if the threshold for class 1 is set to 50% then the output of the model is labeled 1. If we set the threshold to, say 70% then the output changes to 0. So playing around with the threshold gives you different outcomes for the ROC curve.
Thank you so much for explaining these ideas with this concise video.
Thanks for taking Time and explaining so well
Thank you. This is much understandable than my textbook
Excellent explanation
Thankyou so much. Brilliant video !
You're most welcome
Great video! Suggestion: Normalize volume to 50% going forward as I really had to crank up the speakers to hear your voice.
I am so grateful.
Thank you.
Good Content
Subscribed right away!!!
Thank you very much!
Thanks! More videos please!
great explanation
Stunning!!
Very well explained. Thank you very much. I just pressed the Subscribe button :)
Great intro
its great simple video will be great to do more videos showing the over fitting and under fitting and other questions that normally been on interviews
Thank you for the positive feedback. I'll do my best.
Thank you for the vid 👍
But what do you mean by "thresholds" at 11:10 ?
Like, what are the thresholds in terms of neural networks, and how can we change them?
Thank you :)
Thank you for the positive feedback :)
The most simple way to think of a threshold is through a simple model, with only 1 feature as an input and two possible classes as output (say 0 and 1). Then when the model is trained it finds the "best" threshold for the input feature. So, for example, if we denote the input feature as a, then the model may learn that if a>=0.5 then the label is 0, otherwise it's 1. In this example the threshold is 0.5.
Neural networks work differently (in most cases) and thus thinking of a threshold may be confusing. For our example from the paragraph above, the output of a neural network will be a vector/list of size 2, where each index is the probability that the output is in that specific class. For example if the output is [0.14, 0.85] then the model "thinks" that there is a 14% chance that the input is from label 0 and a 85% chance its from label 1. If our neural network had only 1 neuron then the 0.5 value from the example above could be incorporated into it.
"How do we change it?" - This really depends on what you want to achieve. If FA are more important than FN, or the other way around then you can change your loss function and the incorporated threshold will change accordingly.
Hope this helps :)
@@ml_dl_explained cool, thanks 👍
Tbh, I didn't really get how can we "vary the thresholds" to further plot ROC or AUC for a neural network.
.
I mean, when a model is trained - we have only a single point at the ROC plot (current state of the model).
But then how can we "change the thresholds" to have multiple points on the plot?
Thank you 😊
You're welcome :)
When we train the model, it outputs probabilities. We can change the threshold of those probabilities to get different labels - for example, if the model's output is [0.34, 0.66], one threshold could be if the threshold for class 1 is set to 50% then the output of the model is labeled 1. If we set the threshold to, say 70% then the output changes to 0.
So playing around with the threshold gives you different outcomes for the ROC curve.
@@ml_dl_explained oh, okay, so ROC and AUC are mostly used for the binary classification?
Yes. exactly
Great video
Thanks!
Brilliant.
Thank you
Thank you so much 💙💙💙💙💙🌌🌌🌌.
Thank you for the positive feedback
Can I have these slides please within respective concern 🙏💓
well explained
Thank you :)
thanks sir.
Awesome!
Thank you! Cheers!
nice
Mertz Vista