Introduction to machine learning and deep learning
0 likes964 views
The document discusses the importance of machine learning (ML) and data analytics in making data-driven decisions, optimizing predictions, and understanding user behavior. It highlights the significance of deep learning for handling large datasets and high-dimensional data and addresses common challenges such as bias, variance, and data quality. Additionally, it provides insights into different ML problems like classification and regression, along with resources for further learning.
Introduction to machine learning and deep learning
- 1. Introduction to Machine Learning & Data Analytics Shishir Choudhary
- 2. Why ML ? What happened with user at meeting today ? What user has bought in past What happened at user house today morning ? What people with similar interest bought What has user recently shown interest in by search ML Prediction System You want to take a good decision • You can’t ask a human to take the decision due to time and dollar cost involved at the scale of problem • You are presently anyway not taking good decisions because decision is presently random , hard coded or complex rules based • You want to reduce cost or verify your own decision E.g. Taking good decision is neither easy, nor perfect, but ML makes it less error prone than simpler old alternatives
- 3. Why Data Analytics ? Understanding your business/industry/users better You want to improve your business, industry or users, e.g. What segments of users you have and how well each segment likes your product. How well are different players in industry doing in different segments of products they well ? Which products are growing fast ? Helps to remove bias or opinions and gives more factual insights. Also called data driven decision making. Typically involves visualisation, clustering and at times regression(ML). Sometimes this involves analysing “big” amount of data typically from world or “small” amount of data from your company itself
- 4. Why deep learning ? • Better results with large amount of data • Very high dimensions in data or time series / sequence data - E.g. Image, Voice, Text, Time series • End to end Vs Part of workflow
- 5. What is ML = Optimisation Problem Given Past Data Parameter that can’t be obtained What user has bought in past Parameter that can’t be obtained What people with similar interest bought What has user recently shown interest in by search What product should be recommended to user ? Imagine 20 inputs Prediction System Weight for decision = 0.2 ? Weight for decision = 0.3 ? Weight for decision = 0.5 ? ML = Optimize the Weights Past Purchase Data
- 6. What is ML = Inductive Reasoning Problem • Guess the function • f(1) = 1 • f(2) = 4 • f(3) = 9 • f(4) = 16 • ?? what is f(x) ? Involves leap of faith. Does training generalise well is the crucial question ?
- 7. Alternative = Deductive Reasoning = Rules Based System Simple attempt at predicting gender by rules • Men => Short Hair • Women => Medium/Long Hair • Arun is a Man => Arun has short hair Quite quickly, rules and their exceptions can become too complex to manage
- 8. Most Popular ML Problem Statements • Classification • Given lot of examples, learn how to predict function(Input) -> Output one of Fixed Set Of Categories • E.g. Spam or Not • Good vs bad investment • Image classification / labelling • Regression • Given lot of examples, learn how to predict function(Input) -> Continuous infinite values function • House Price estimation • Pricing a stock • Time for food to arrive for food delivery app
- 9. Classification or Regression ? $100-120k $120-140k $140-160k $160-180k 100 120 140 160 80 sqm 100 sqm 120 sqm 140 sqm House Pricing - Classification Class 1 Class 2 Class 3 Class 4 House Pricing - Regression Size -> Size ->
- 10. Crucial challenges • Bias Error & Variance Error & Irreducible Error ML Models try to balance between bias and variance. Some models by default have more bias and some have by default more variance. Some some parameters their default leaning can be adjusted. DL can reduce both bias and variance error with huge amount of data and high number of features/parameters
- 11. Crucial Challenges • Curse of dimensionality • Amount of training data needed increases exponentially with more features used for prediction • Hence dimensionality reduction techniques • Bad and missing data • Lot of real world data often is bad quality, with either wrong or missing values. Thus good data becomes quite useful asset today. • Feature Engineering • In ML you need to select right features to use for prediction. This is a tough problem and partly solved by domain knowledge and partly by data analysis
- 12. ML vs DL 0 25 50 75 100 Untitled 1 ML DL Small DL Large Larger Training Data Accuracy At lower quantum of training data, ML , Small DL models and Large DL models all perform quite similarly. At larger quantum DL is better able to utilise the additional training data. Latest Image Recognition DL Models perform better than human error rates.
- 13. Models ! Models are fancy words for mathematical functions which are used to generalise from training data
- 14. Prediction Optimisation Most learning algorithm work in some sort of derivative of following steps 1. Based on inputs and initial model parameters, make a prediction 2. Check what was actual answer and calculate error 3. Try to minimise the error by adjusting the models’s parameters in direction which reduces error 4. Repeat till convergence (error % stops reducing and is low)
- 15. Regression Takes simplifying assumption that prediction is linear (or binomial or trinomial) relation to input(s) E.g. Height of plant by age , time required to deliver food in food app
- 16. Decision Trees
- 17. Perceptron
- 18. Neural Network
- 19. Good resources • Udacity (completely) Free Machine Learning Course • Josh Gordon’s Youtube Videos on Introduction to ML • Libraries • Scikit Learn in Python - Machine Learning • Tensorflow - Deep Learning • Gensim - Natural Language Processing • http://learney.sg/#machine learning • http://learney.sg/#data science • http://learney.sg/#deep learning • http://learney.sg/#natural language processing
- 20. Summary & Q &A