We use www.menti.com to communicate with teacher
Steps of ML application
1. Select problem
2. Get data
How long do you spend on collecting data?
It’s recommend that quickly collect data at the firt time as we start the project, so we can modify the design of the model or collect more data by training the model.
Keep clear notes on reperlment line(?).
How would you collecting data?
3. Design model
4. Train model
5. Test model
6. Deploy
Edge devices
audio —> NN —-> 0/1
but their might be large NN, it waste times.
audio —> VAD —> NN —> 0/1
VAD is checking if the voice actually detective.
option 1 will be more simpler and faster to option 2. It maybe have many errors or noises in the result of option 1, but the following large NN will take care of it, we don’t have to worry about it.
In this case, option 2 will be a proper choice for Non-ML just detect for loud or seilence, it can not distinguish the anscent.
7. Maintain
Note that the steps might go backward or step into cycling when the step are not satisfied.
2018/10/17
Attacking a network with adversarial examples
GOAL: Given a network pretained on ImageNet, find an input image that is a cat but will be classify as an iguana.
1. Rephrasing what we want
2. Defining the loss function
3. Optimize the imagine
The pic looks ‘real’ to human might be fake of the ML result, the result mechine classified into iguana might looks like a cat to human, we have to avoid it.
Sol.
Create a SafetyNet, classify all images at the first time as ‘T’ or ‘F’.
Train on correctly labelled adversarial examples.
Adversarial training
Linear Regression
g = Wx + b
X = [x1; x2; x3; x4; x5; x6]
number of hidden layer = 1
w = [1, 3, -1, 2, 2, 3]
b = 0
Given X = [1; -1; 2; 0; 3; -2]
yhat = 1 - 3 - 2 + 0 + 6 - 6 = -4
xStar = x + epsilon wT
yStar = w xStar = wx + epsilon w wT
where w * wT = ||w||^2
let epsilon = 0.2,
xStar = x + epsilon wT
= [1 + 0.2; -1 + 0.6; 2 - 0.2; 0 + 0.6; 3 + 2.6; -2 + 0.9]
= [1.2; -0.4; 1.5; 0.6; 5.6; -1.1]
yStar = w xStar = …
epsilon is not growing with y dimension, but it impacts the complexity of yStar calculating.