Thank you so much for this explanation. I sure hope there will come more and we dive in deeper. I especialy love that you take those anthomophised terms and bring them back down to the actual math terms.
Nice animations, but too many unexplained or hand-wavy manipulations. 22:23 says could just multiply cost by inputs. How does this knowledge fit the rest of the video or is transferrable to other networks? This manipulation tells me it's no longer MAE but MSE as cost. If true that's such an important change that should be emphasized since much of a network's utility and ability to be trained is based on what loss it works with. 19:37 says f = wx+b. In this video, f had 2 meanings, as a function, and as the output. That is immensely confusing to beginners. 18:39 dc/df = sign(t-p). Usually we expect f to appear in the derivative expression (unless it's a constant), here f turned into t and p, it was confusing. This is related to previous point that f can both be a function and p in this video. 18:12 the yellow box shows f(w*x+b). Usually we would expect f(x), instead of applying f on w*x+b as this notation implies. 21:10 random addition of abs out of nowhere. Why wasn't it added the first time this equation appeared? 22:26 "because the perceptron updates w and b in the direction that minimizes the error" is a general statement that applies everywhere. I don't see how it explains why the gradient equation changed
Thanks for taking the time on such an elaborate feedback!! You have some valid points! I tried to keep the video below 30 mins, but when I finished shooting it it was 1h:10m long so some things had to go !!! oh well, I will try to do better on the next one mate! I am still knew at this youtube'ing haha :)
So, to be clear-- error, cost and loss are all the same metric? Would this be true even in a framework like TensorFlow? Addtionally, what are some good resources you can reccomend for learning the very basics of Calculus? To be honest, my program never required a calculus course to be taken, and my background on calculus is pretty weak, and it made the end of the video a bit confusing for me.
That's right mate! Some people in the industry refer to error as the difference between predicted values and the target value, loss as a error for a single traiining example and cost as the average loss/error as I shown in the video. But to be honest mate, that's all semantics to describe the same thing! In the end, and especially when I make the next videos, what matters is the average cost and how to backpropagate the error/loss/cost. Now in the context of a framework like tensor flow or pytorch, they might refer to it as just one of those words (depending on the framework) but they all mean the same. Same way as the word "dense" means the number of neurons. But some frameworks just refer to it as "dense" or "neurons"
Sorry I missed the calculus part of your message. There are a myriad of good resources out there mate. Khan academy is my favourite, but there is a really nice lady on youtube called NancyPi who also explains it beautifully! There is also a really nice course on udemy called "essential calculus for neural networks" which is quite good as well. But don't buy the 50 quid version, wait for a sale and you can get it for like a tenner. Hope this helps and thanks for watching! :)
![The moment we stopped understanding AI [AlexNet]](http://i.ytimg.com/vi/UZDiGooFs54/mqdefault.jpg)
You have a excelent teaching skills, please continue with the ML and mathematics topics
Thank you very much mate! :) Appreciate the nice words!
Thank you so much for this explanation. I sure hope there will come more and we dive in deeper. I especialy love that you take those anthomophised terms and bring them back down to the actual math terms.
think I'm a bad zane, need to watch again this video to understand :) :)
heheh, as always, thanks again for the support mate!
Awesome,btw,I'm Zane 😄🙌🏼🙌🏼🙌🏼❤
lool :D
I guess in MAE we should get the absolute difference using abs function.
costs = [abs(t - p) for p, t in zip(pred, targets)]
Came from reddit, function approximation 🤯🤯 nice vid 👍👍👍
thanks mate!! :)
😊great video
❤
Nice animations, but too many unexplained or hand-wavy manipulations.
22:23 says could just multiply cost by inputs. How does this knowledge fit the rest of the video or is transferrable to other networks? This manipulation tells me it's no longer MAE but MSE as cost. If true that's such an important change that should be emphasized since much of a network's utility and ability to be trained is based on what loss it works with.
19:37 says f = wx+b. In this video, f had 2 meanings, as a function, and as the output. That is immensely confusing to beginners.
18:39 dc/df = sign(t-p). Usually we expect f to appear in the derivative expression (unless it's a constant), here f turned into t and p, it was confusing. This is related to previous point that f can both be a function and p in this video.
18:12 the yellow box shows f(w*x+b). Usually we would expect f(x), instead of applying f on w*x+b as this notation implies.
21:10 random addition of abs out of nowhere. Why wasn't it added the first time this equation appeared?
22:26 "because the perceptron updates w and b in the direction that minimizes the error" is a general statement that applies everywhere. I don't see how it explains why the gradient equation changed
Thanks for taking the time on such an elaborate feedback!! You have some valid points! I tried to keep the video below 30 mins, but when I finished shooting it it was 1h:10m long so some things had to go !!! oh well, I will try to do better on the next one mate! I am still knew at this youtube'ing haha :)
So, to be clear-- error, cost and loss are all the same metric? Would this be true even in a framework like TensorFlow?
Addtionally, what are some good resources you can reccomend for learning the very basics of Calculus? To be honest, my program never required a calculus course to be taken, and my background on calculus is pretty weak, and it made the end of the video a bit confusing for me.
That's right mate! Some people in the industry refer to error as the difference between predicted values and the target value, loss as a error for a single traiining example and cost as the average loss/error as I shown in the video. But to be honest mate, that's all semantics to describe the same thing! In the end, and especially when I make the next videos, what matters is the average cost and how to backpropagate the error/loss/cost. Now in the context of a framework like tensor flow or pytorch, they might refer to it as just one of those words (depending on the framework) but they all mean the same. Same way as the word "dense" means the number of neurons. But some frameworks just refer to it as "dense" or "neurons"
Sorry I missed the calculus part of your message. There are a myriad of good resources out there mate. Khan academy is my favourite, but there is a really nice lady on youtube called NancyPi who also explains it beautifully! There is also a really nice course on udemy called "essential calculus for neural networks" which is quite good as well. But don't buy the 50 quid version, wait for a sale and you can get it for like a tenner. Hope this helps and thanks for watching! :)