Things that you should know in ml

This is presentation is to understand the basic terms that are used in ML. I covered the most frequently used topics.
#UFest18 #IndiaMLCC #UFestIndia2018

Transcript

1. Krunal Kapadiya
   Engineer @
   @krunal3kapadiya
   #IndiaMLCC #UFest18
   1
   

   View Slide

2. View Slide

3. Input Output
   Magic
   What is Machine Learning?
   3
   

   View Slide

4. Types of machine learning
   Based on Human supervision
   Supervised ML
   Unsupervised ML
   Semi supervised
   Reinforcement learning
   4
   

   View Slide

5. Types of machine learning
   Based on learning
   Online learning
   Offline learning
   Based on Data patterns
   Instance based
   Model based
   5
   

   View Slide

6. Measures of central tendency
   What is Mean?
   - Average value from dataset
   What is Median?
   - Middle value of dataset
   What is Mode?
   - Repeated value of dataset
   http://bit.ly/MeasureCentralTendency
   6
   

   View Slide

7. Skewness
   Remember
   Negative Skew: Mean is less than mode
   Positive Skew: Mean is greater than mode
   7
   

   View Slide

8. Data scientist ≠ Data Engineer
   Machine learning ≠ Data scientist
   8
   

   View Slide

9. Training set and Testing set
   9
   

   View Slide

10. View Slide

11. View Slide

12. View Slide

13. View Slide

14. “A”
    - Auto regression - Time series regression model
    - Activation Function
    - Sigmoid - Use in binary classification
    - ReLU - Helps in hidden layers
    - Softmax - Mostly used in multiclass classification
    - A/B Testing - Which technique perform better
    - Accuracy - correctly predicted values
    14
    

    View Slide

15. When you learn ML in 24 hours
    15
    

    View Slide

16. “B”
    - Bagging - multiple models and final prediction combining all predictions
    - Box plot - displays range of variations in data
    - Backpropagation - update weights reduce errors
    - Batches - small chunks and splitted data
    - Batch normalization - improve performance and stability of DNN
    16
    

    View Slide

17. “C”
    - CNN - Convolutional Neural network
    - Classification - Labels are known
    - Cost function - cost function minimum, models accuracy best
    - Confusion Matrix - displays performance of the model
    17
    

    View Slide

18. “D”
    - Dropout - Hidden layer dropped to prevent overfitting
    - Data Augmentation
    18
    

    View Slide

19. “E”
    - Eager execution - operations runs immediately, waiting for graph execution
    - Epochs - single training iteration
    - Early stopping - prevent overfitting
    “F”
    - Forward propagation - only one way input to output, no backward
    19
    

    View Slide

20. “G”
    - Gradient Descendant
    - Batch GD
    - Stochastic Gradient Descendant
    - Mini batch GD
    20
    

    View Slide

21. “G”eneralization (a.k.a out of sample error)
    - Measure of accuracy for previously unseen data
    - Difference between expected and proven error
    - Mostly occurs in deep learning model, training sets working fine, but not fitting in
    real data
    21
    

    View Slide

22. “H”
    - Hyperparameters - values set before training model, e.g. batch size, number of tree
    - Histogram - use to determine skewness
    “I”
    - Imputation - wrangling data, filling missing values
    22
    

    View Slide

23. “L”
    - Learning rate - amount of minimizing in cost function
    - LSTM - building units in RNN, speech pred, rhythm learning
    “M”
    - MLP (Multilevel perception) - aka fully connected layers
    23
    

    View Slide

24. “N”
    - Numpy
    - Linspace
    - Random
    - Array
    - Arange
    “O”
    - Outliers - value that far away from dataset pattern
    24
    

    View Slide

25. “P”
    - Pandas
    - Dataframes
    - Series
    - Pooling - use to reduce parameters and prevent overfitting
    “R”
    - Regression - predicting values, typically in floating points
    25
    

    View Slide

26. Pooling
    26
    

    View Slide

27. Validating model
    Confusion Matrix
    - If false negatives are ok, requires high
    precision, e.g. Spam filter
    - If false positives are ok, requires high
    recall, e.g. Medical Diagnosis
    Precision Recall
    F-1 Score Accuracy
    Accuracy = Ratio of correctly classified
    points / total points
    27
    

    View Slide

28. Let’s Go For It
    1. Look at the dataset
    2. Write down columns and it’s correlation
    3. Make questions derived from the dataset
    4. Explanatory Analysis with visualization
    5. Frame problem
    6. Create solution by creating model
    28
    

    View Slide

29. Explanatory Analysis
    - Look at the rows and tables
    - Find correlated columns
    - Display it in charts
    - Give summary based on graph
    29
    

    View Slide

30. TMDB Notebook (Dataset)
    30
    

    View Slide

31. Reference
    https://www.analyticsvidhya.com/blog/2017/05/25-must-know-terms-concepts-for-begi
    nners-in-deep-learning/
    https://ml-cheatsheet.readthedocs.io
    31
    

    View Slide

32. View Slide

33. Thank You
    33
    https://krunal3kapadiya.app/
    @krunal3kapadiya
    #IndiaMLCC #Ufest18
    

    View Slide