Deep-Learning-Course/SecondSaturday/Notes.md

1. flowchart application -> ios mobile find symbol through gestures.
2. matrimony portal profiles -> finding matching profiles
3. swipe input keyboard for indian languages.
4. mnist hand-written digit database -> build application for recognizing full phone numbers(10 digit).
5. live cricket feed -> generating highlights of the match.
6. designing a chatbot for getting cricket scores.


# general approach to machine-learning
1. model, objective and learning algo
3. ml is technique for learning from examples
    -> recommending smart-phone to a friend. price, branch, cam quality, screen size, processing speed. ==> model
    -> objective ( why is he buying the phone) should be able to be boiled-down to a single number.
    -> learning algo
    binary features -> camera 0|1
                    -> screen small|large
                    -> battery good|bad
                    -> memory high|low
                    -> processing fast|slow
                    -> audio good|bad

prior probability -> probability of an event occuring without any knowledge of conditions

P(A)*P(B|A) = P(B)*P(A/B) = P(A,B)

P(+ve|[x1,x2,x3...xn]) = P([x1,x2...xn]|+ve)*P(+ve)/P([x1,x2...xn])
                       = P(x1|+ve)*P(x2|+ve)*P(x3|+ve)*P(+ve)/P([x1,x2...xn])
                       = Pi (i=1 to n) P(xi|+ve) * P(+ve)(Class variable)/ Sum (C=+ve to -ve) (P([x1,x2....xn],C))
                       = Pi (i=1 to n) P(xi|+ve) * P(+ve)(Class variable) / (( Pi (i=1 to n ) P(xi|+ve)*P(+ve))+..+ Pi (i=1 to n) P(xi|-ve)*P(-ve))

P(X,Y) = Sum (y=y1...yn) P(X,Y=y)

W2 = P(+ve|xi=1) (human approach) (Naive bayes) (naively thinking that all features are independent)


Regression : output contains real values -> (predicting the position of joints of a body given an image of a person)
Classification : output classifies to discrete set of classes
      -> predicting the posture of a person (sitting,walking,standing,running) given an image of a person
(Numerical/Categorical)

Representation Learning: embedding

Deeplearning is all about hierarchical representation learning.

Metric Learning: distance( of facial features)/similarity(of fashion apparels)/relevance( of search document)
Structured Output(Models): auto-corrects dependent outputs based on output on the upper hierarchy outputs.

Types of input:
Bag of features,bag of words: ( finding whether a feature is present/not without caring where the feature occurs in the input)
eg: Using unsupervized learning to convert the input to a given set of classes(clusters) and use them as bag of features.

Spatial data(sequential data): if there local dependencies use CNN(convolutional nn) if there are near past dependencies in the data use RNN(Recurrent NN-LSTM)
eg: stock market temporal data / speech data/ image data

Non-Parametric models : k-NN(K-nearest neighbor), Decision Trees, Random Forests (independent of parameters)
  -> very inaccurate because doesn't know much about the data
Parametric Models: based on fixed set of parameters,SVM
  -> more accurate coz the knows more about the parameters from the data


Types of Learning:
  supervized learning -> labeled data
  unsupervized learning -> unlabeled data
    exercise: *take 3s from mnist data *create a gmm model with them and *cluster them with 5/3/10 gaussians.
              *take all images and cluster them to 10 gaussians.
  semi-supervized learning -> combination of supervized and unsupervized models

  Auto Encoder:
    finding a low dimensional representation of a high dimensional data.
    eg. image of 200x200 pixels create a fingerprint of image of 128 dimensions.
       exercise: use the 128 dimensional data to reconstruct the 200x200 image(using inverse of the model).

  Reinforcement learning:
    eg:playing chess -> using the final result of the game to assign weights/score for moves that were made upto the final result. and training the model to predict based on those scores.