Lesson 3
Note: content starts at 45min *
rectified linear unit(relu)- max(0,sum) in excel
Takes the on the positive unit which gives it the horizontal edges
max pull
tensor - is an array with more dimensions so we can stack kernals together
they are multi dimensional
kernals/filters are being used on multiple the convulsions that are created before it
Max Pooling- using 2 by 2 pooling we would have the half resolution(replace four (numbers) cells with the biggest number)
Fully connected layer - sum product of every activation by each weight(using weight matrix). Don't use these much anymore
- Activation function - eg: softmax. Takes in 1 number spits out another number, nonlinearity find a better definition
Softmax binary classification (single label)
insert image of softmax eq
Multi label Image classification
Starts lesson3 3 1:20
Satellite imagery can have things many labels like weather, agriculture, primary (rain forest) water
Sigmoid use for multi variable classification(multiple label)
$$s(x)= \frac{e^x}{e^x+1}$$
learn.summary()
shows us our model
Structure of Data
unstructured -all the data is the same thing eg pictures, language, sounds waves
structured - database type of structure
Structured data
Lesson 3- roossman starts at 2:03
fastai.structred
for structured data can be used by itself without any fastai packages.
fast.ai columnar
lets us work with columnar data
Check out
- keras lesson 1 to see lesson one written using keras
- multi dimensional array
- convolution slide
- logs rules and e