# HTML HW4 Grading Policy

## P3
- 20 Points: Correct solution. Using $w=V\tilde{w}$ or taking gradient and compare are common reasonable solution.
- 15 Points: Almost correct: there are minor errors in the derivation. e.g.
- power of a vector
- a vector is divided by another vector
- vector dimension mismatch
- did not assume $U$ is symmetric and gave the derivative the $U$ term to be $2U$ (should be $U+U^T$).
- 10 Points: Partially correct: some derivation is reasonable, but the result is wrong or has major errors, or the derivation is incomplete. e.g.
- you tried to compare two optimization formula seriously.
- compute gradient
- 5 Points: The derivation is incorrect but you demonstrated reasonable efforts.
- 0 Points: No derivation is given or unrelated.

## P6
**Correct answer: (b)**

- 20 points
Clear and correct solutions.

- 15 points
There's a minor issue in your solutions, e.g.
- Miss the identity after $\frac{2\lambda}{N}$.
- do matrix devision

* 10 points
A major issue, e.g.,
- Calculation mistakes which leads to a wrong answer

* 0 points
* Unrelated explanations
* Without explanations

## P9
**Correct answer: (b)**
- 20 points: Correct
- Any logical derivation.
- 15 points: Minor issues
+ Lack of clarity in some steps.
+ Missing $\mathbb{E}[\cdot]$ throughout the derivation.
+ Took $\bar{y}$ out from $\mathbb{E}[y_n\bar{y}]$. (The correct way of solving $\mathbb{E}[y_n\bar{y}]$ is by using the law $\mathbb{E}[XY]=\mathbb{E}[X]\mathbb{E}[Y]$ if $X$ is independent of $Y$)
- 10 points: Major issues
+ Assumed $y_n$ are drawn from a certain distribution.
+ Incorrect derivation of $\mathbb{E}[\bar{y}^2]$.

- 5 points: Showed some reasonable effort
+ Fundamentally wrong.
+ Assumed $\bar{y}=0$.

- 0 points: Did not show any steps
+ No pages selected.
## P10
**Correct answer: (a)**

- 20 points: correct solution
+ Basically, if you convince me you can derive the correct $E_{in}$ and probability (showing how it would result in an error), you will get a total score

- 15 points: minor issue
+ Writing $+-+-$, $-+-+$ or OXOX, XOXO without explaining what they represent (the order of the vectors)
- Simply saying there are two combinations without any explanations (At least say we learned 2D perceptron from before...)

* 10 points: major issue
+ Correct probability but wrong $E_{in}$
+ Correct $E_{in}$ but wrong probability

* 0 points
+ Totally wrong or lack convincing calculation