Timeline for How to resolve the perceptron dilemma for binary classification?
Current License: CC BY-SA 4.0
9 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| S Jun 10, 2021 at 8:04 | history | bounty ended | CommunityBot | ||
| S Jun 10, 2021 at 8:04 | history | notice removed | CommunityBot | ||
| Jun 2, 2021 at 23:38 | answer | added | Lucas Prates | timeline score: 1 | |
| S Jun 2, 2021 at 6:16 | history | bounty started | Olórin | ||
| S Jun 2, 2021 at 6:16 | history | notice added | Olórin | Draw attention | |
| May 25, 2021 at 18:55 | history | edited | Olórin | CC BY-SA 4.0 |
added 1 character in body
|
| May 25, 2021 at 2:23 | comment | added | microhaus | On that basis, my instinct would be that there is no dilemma at all. The intuitive reason why I hold that belief is that "squishing the data" together more closely via normalisation will result in the distance of the closest point to the optimal decision boundary, i.e. the margin $\gamma$, also reducing by exactly the same factor. Meaning that no relative distances are altered - you've just 'zoomed outwards'. However, a full answer to substantiate those intuitive reasons would require a formal derivation. | |
| May 25, 2021 at 1:51 | comment | added | microhaus | 1. In my view, the 'dilemma' only follows if you grant that the net effect of normalisation is for the bound $R^2 / \gamma^2$ to increase. 2. For this net increase in the bound to occur, it would have to be the case that normalisation would entail a reduction in the margin $\gamma$ that is greater than the reduction in the bound on the Euclidean norm of the data $R$. 3. To me, it is unclear why point 2 would be the case, rather than for the bound $R^2 / \gamma^2$ to be unchanged from normalisation because the reduction in $R$ is exactly equal to a reduction in the margin $\gamma$. | |
| May 24, 2021 at 23:45 | history | asked | Olórin | CC BY-SA 4.0 |