2017-09-07
HW01: Winnow Algorithm for Threshold Functions.
Maybe this is a dumb question, or maybe I am giving away the game. If it is the latter, feel free to delete this question.
.
However, if I understood the Winnow algorithm correctly, since `\alpha > 1`, by the definition of unit spherical, and random.random() always returning a positive number, it seems that the product of `w_i` and `x_i` could never be negative. Hence, for linear thresholds like the one you gave (or many others), it could never properly handle the negative (e.g., `-5`) weight.
.
Is that a correct understanding, or did I overlook something obvious that I will regret making this post?
Maybe this is a dumb question, or maybe I am giving away the game. If it is the latter, feel free to delete this question.
.
However, if I understood the Winnow algorithm correctly, since @BT@\alpha > 1@BT@, by the definition of unit spherical, and random.random() always returning a positive number, it seems that the product of @BT@w_i@BT@ and @BT@x_i@BT@ could never be negative. Hence, for linear thresholds like the one you gave (or many others), it could never properly handle the negative (e.g., @BT@-5@BT@) weight.
.
Is that a correct understanding, or did I overlook something obvious that I will regret making this post?
Yes, that's correct. You can do pre-transformations of the data to handle this kind of situation, but you don't need to do that for this homework.
Yes, that's correct. You can do pre-transformations of the data to handle this kind of situation, but you don't need to do that for this homework.
2017-09-08
Thank you for the clarification. I went back and checked the slides, and I am not sure if the note concerning Winnow always being positive is new or if I missed it originally. However, I did notice there is a typo on that slide:
.
"Given the set-up on initial weights and α, at any training step using Winnow, the weights will always eb positive."
(Edited: 2017-09-08) Thank you for the clarification. I went back and checked the slides, and I am not sure if the note concerning Winnow always being positive is new or if I missed it originally. However, I did notice there is a typo on that slide:
.
"Given the set-up on initial weights and α, at any training step using Winnow, the weights will always <u>'''eb'''</u> positive."
2017-09-09
Fixed. The bullets were new. Since this is my first time teaching this class, I am going over my slides after lecture as well attempting to improve them.
Fixed. The bullets were new. Since this is my first time teaching this class, I am going over my slides after lecture as well attempting to improve them.
(c) 2024 Yioop - PHP Search Engine