recurrent neural nets Flashcards
1
Q
What are the applications of recurrent neural nets?
A
- Speech recognition
- Text recognition
- Stock prices
- Sequential data
2
Q
Vanilla supervised learning and Structured Input
A
- vanilla supervised models (feedforward nets) assume no input structure. These models are not good for analyzing organized info like sentences
- Doesn’t care about the order of the info
3
Q
ways to model ordered sequences?
A
- recursive functions. we want given a sequence input model it as a recursive function
- First approach is to model the recursive function as a linear function and try to fit the weighs. Do this with minimized least squares sum. This is a linear regression
4
Q
Using keras for creating a recursive linear funtion
A
- Use the dense layer with activation linear
5
Q
logistic recursive sequence
A
s1=α
s2 =ws1 −ws1^2
6
Q
What is windowing?
A
- The process of processing the sequence for training
- Obtaining the input - output pairs
7
Q
Flaws with FNN approach
A
- The probabilistic interpretation of least squares is that input/output pairs are independent and identically distributed. This is not good because it means data don’t have structure
- With iind changing a previous value the forward values don’t change
- We make things independent the opposite we want