Deeplearning,transfer learning,broader view

Question 1

Kernel

Answer 1

Kernel function
Non-linear phenomena can be linearized using
a kernel function
* Non-linear mapping + linear model
* Easy to implement / relies on simple models
* Kernel is often application-dependent, hard to
generalize

Question 2

DL – advantages

Answer 2

Handcrafted features might be incomplete
and require a lot of human work for designing
and validation
* Learned features are adapted to the input
data
* Learned representations work for both
supervised and non-supervised applications
* Data representation and classification into one
single network – end-to-end learning.

Question 3

DL Actiavtion functions:

Answer 3

Sigmoid neurons saturate and
kill gradients

Tanh function:
* Saturation: similar to
sigmoid
* Output is zero-centered
* Tanh is a scaled sigmoid:

RELU
Faster training thanks to its
non-saturating nature
– Prevents the gradient
vanishing problem
* Lighter computation

Question 4

Fully Connected feedforwardNetwork

Answer 4

Several layers
No loops
High nr of connections
Usage:Classification

Question 5

FROM NN TO CNN

Answer 5

1-Local connectivity:
only neighbors are connected. lower number of weigts to train

2-Shared weights.
many units share parameters but on diferent input windows.

3-Multi feature maps
Compute different function on same input.

4-Subsampling(pooling)
Reduces resolution
reduce computional complexity
adds some deformation to invariance
max pooling fast effixient and strong.

Question 6

CNN regularization

Answer 6

1-Dropout
Randomply drop units with fixed proability
2-Early Stopping.
Stop when validation error is not improved
3.Preprocessing
add noise rotation flip etx data augmentation.
4.Batch processing because gradient is slow.

Question 7

Video Sequences

Answer 7

-Optical Flow
2D motion of image between consecutive images.
3D motion to 2 D.

-Motion flow: actual 2d motion of points.

Question 8

Beyond Natural Lighting

Answer 8

Lambertian material: Perfect difuse

Question 9

EM spectrum

Answer 9

much wider than visible specturm

Multispectral imaging
-wider portion of em specturm
-multiple bands

Hyperspectral
-wider portion of em spectrum
-fine grained sample
-Composed of a huge number of channels
-Measure “chemical”-related properties of
materials
-Enable multiple analysis to be performed with one
single image