Getting started with Machine Learning
1 like405 views
The document provides an overview of machine learning, including definitions, types of machine learning algorithms, and the machine learning process. It defines machine learning as using algorithms to learn from data and make predictions. The main types discussed are supervised learning (classification, regression), unsupervised learning (clustering, association rules), and deep learning using neural networks. The machine learning process involves gathering data, feature engineering, splitting data into training/test sets, selecting an algorithm, training a model, validating it on a validation set, and testing it on a held-out test set. Key enablers of machine learning like large datasets and computing power are also mentioned.
Getting started with Machine Learning
- 1. Getting started with Machine Learning Gaurav Bhalotia
- 2. Machine learning ● Processing Data, understanding data, reacting intelligently to data ● Predicting the future based on the past Traditional CS Machine Learning
- 3. ML vs AI? AI is a computer program that does something smart. It can be a pile of if-then statements or a complex statistical model. Not all AI counts as machine learning. For example, symbolic logic (rules engines, expert systems and knowledge graphs). Trie based autosuggest are e.g. Computer programs (algorithms) to parse data, learn from it, and then make a determination or prediction about something in the world. Teaching computers to learn to perform tasks from past experiences (data) A technique for implementing machine learning. Attempts to model the mind rather than the world
- 4. What is learning? Do the actual maths/calculations (training) Figure out a mathematical formulation (model) Identify the real world problem we are solving. (the question) Find the relevant data, gather enough so that it captures all nuances 1 2 3 4
- 5. Machine Learning has come of age Multiple Real World Deployments Speech recognition (Siri, Alexa, … ) Image recognition (self driving cars, FB tagging … ) Spam detection Recommendation Engines (amazon, netflix, … ) Language Translation Data1 Modelling Capability 2 Compute Power 3 Programming Frameworks 4
- 6. What can Machine Learning/ AI do? Anything humans can do in < 1s (Andrew Ng @ Baidu) Predicting the next in a sequence of events We are very very far away from AI becoming sentient
- 7. Structuring this a bit
- 8. Model Input Learning Gather data: Identify data context that represents the problem space. Ensure sufficient coverage for various nuances. Feature engineering: If shallow learner, identify specific features of the input that may help in disambiguating the correct answer. Data Preparation: Split data into training, testing and validation sets. Assign human labels to training data if fitting to be done against known outputs. Learning Type: Identify the class of algorithms to be used - supervised, unsupervised, etc. Algorithm Selection: For the chosen class, pick one that may work best for problem at hand - e.g. classification neural net vs SVM vs naive bayes Parameter Identification: Figure out the parameters or a set of parameters that you want to try for your algorithm during the learning process. E.g. #cluster for k-means Train: Fitting the input data to the model we are working with. Gradient descent is a well known tool for this. Validate: While the model is being trained on training data, we update the model only if it continues to do well on an independent validation set. Test: Test how well the model is able to generalize the learned behaviour by applying on an independent data set not seen before.
- 9. ML Algorithm Types (sample)
- 10. Supervised Learning Supervised learning is the machine learning task of inferring a function (f) from labeled training data ● We have data with already attached labels (e.g. we know spam/non spam difference) ● We predict a pattern that fits the data ● We apply this pattern to new data and predict Goal: from the database (learning sample), find a function f of the inputs that approximates at best the output Discrete/Symbolic output ⇒ classification problem Continuous/Numerical output ⇒ regression problem
- 11. Unsupervised Learning In machine learning, the problem of unsupervised learning is that of trying to find hidden structure in unlabeled data ● We have some (a lot) data that we cannot make much sense of ● We use different algorithms to see if a pattern emerges
- 12. Classification Predicting class labels (two or more) from input data E.g. Spam detection Cancer identification
- 13. Regression Predicting values based on input signals House prices predict Predicting stock prices Predicting LTV
- 14. Clustering or Grouping Grouping data based on certain attributes (distance function) Placing cell towers Identifying patrol vans or emergency clinics Customer segmentation
- 15. Association Analysis Similarity between entities through explicit/implicit associations Supermarket shelf planning Recommendation engines (netflix, amazon)
- 16. Dimensionality Reduction Topic Modelling (news), e.g. LDA Improving performance of ML algorithms
- 17. Sequence Analysis (HMM/RNN) Identify sequential patterns in data Language modelling (speech recognition, language translation) Activity modelling (fitbit)
- 19. Input and Data Preparation
- 20. Feature Engineering ● Examples : bag of words, phonemes ● Using domain knowledge of the data to create features that make machine learning algorithms work ○ both quantity and quality matter ● "Applied machine learning" is basically feature engineering (Andrew Ng). This is what differentiates a expert from novice ○ Is getting replaced with model capability engineering with advent of DL ● The data preparation stage will involve scaling, centering and transforming data to get good features ○ Features should be comparable and combinable
- 21. Data Preparation and Training Training Set: this data set is used to adjust the parameters of the model Validation Set: This data set is used to minimize overfitting. If the accuracy over the training data set increases, but the accuracy over validation data set stays the same or decreases, then we're overfitting and should stop training. Testing Set: this data set is used only for testing the final solution in order to confirm the model performance.
- 22. Picking Training/Test data ● Usually split the input into 60|20|20 or 70|20|10 ● The test set should be similar to train set, don’t train on an entirely different data ● All data should be representative of data that the model will see in future
- 23. Learning
- 24. Model Fitting 1. Start with a seed values for model parameters 2. Identify a loss function which is deviation from input data 3. Fit the model to input data by optimising for a loss function (e.g. gradient descent) 4. Iterate till we converge to a minima
- 25. Nuances and Pitfalls ● Validation set is important ● Monitoring and measuring the performance of model fitting over epochs is important ● The sweet spot is the point just before the error on the test dataset starts to increase ● Use regularisation techniques (dropouts, sampling) The goal of a good machine learning model is to generalize well from the training data to any data from the problem domain.
- 26. ML frameworks
- 27. Deep Learning
- 28. Deep Learning was inspired by how human brain works It’s a massive simplification, but we keep on finding more and more inspirations The brain is way too complex and diverse, but it is still fair to call these biologically inspired Arranged in layers, networks of these neurons are effective at learning from pre-labelled samples (full supervised)
- 29. Amount of data Performance Turning Point in ML Has given us the ability to get to 99% accuracy on so many problems, a game changer
- 30. Why is deep learning taking off now? Scale of data scale of computation At small data sizes, feature engineering wins 1 2
- 31. How does a neural network work? The simplest unit is a perceptron Has ability to model any general function Activation function added to give it non-linearity Are arranged in layers to model more complex functions
- 32. Activation Functions ● Introduced to add non-linearity so that neural network can model any generalised function ● Modelled after biological neuron ● ReLU (rectified linear unit) is a later discovery and has better training performance.
- 33. Network Design ● Hidden layers do implicit feature engineering. ● Softmax used in output layer for classification problems ● The number and size of layers is part of capability engineering ● All weights and biases (parameters) provide degrees of freedom that can fit complex functions
- 34. How does learning happen? 1. Similar to before, start with seed values for weights and biases 2. Compute Loss wrt desired output 3. Find gradient of each parameter wrt the total loss 4. Adjust the parameters proportional to the gradient as part of learning step 5. Iterate
- 35. Training: Optimisation of loss Generalised backpropagation through the layers of the neural network was a breakthrough (Geoff Hinton) Allowed us to train random multi layered networks. Another breakthrough was using massive amounts of data to train computer vision and speech recognition models. (Andrew Ng, Google) Allowed us to achieve accuracies of > 99%
- 36. Backpropagation Our goal with backpropagation is to update each of the weights in the network so that they cause the actual output to be closer the target output, thereby minimizing the error for each output neuron and the network as a whole. We’re going to work with a single training set: We start with random weights and propagate them forward
- 37. Given inputs 0.05 and 0.10, we want the neural network to output 0.01 and 0.99.
- 38. Forward Pass Find what the neural network currently predicts given the weights and biases above and inputs of 0.05 and 0.10, we do this by feeding the inputs forward through the network 1) We figure out the total net input to each hidden layer neuron, 2) squash the total net input using an activation function (here we use the logistic function) 3) repeat the process with the output layer neurons. 4) Finally calculate the error at the output against the expected target
- 39. Backward Pass : Output Layer 1. Find how much a change in each weight affects the total error (partial derivative or gradient wrt weight) 2. To decrease the error, we then subtract this value from the current weight (multiplied by learning rate - here we use 0.5) 3. We perform the actual updates in the neural network after we have the new weights leading into the hidden layer neurons Update each of the weights in the network so that they cause the actual output to be closer the target output : Thereby minimizing the error for each output neuron and the network as a whole.
- 40. Backward Pass : Hidden Layer 1. Find how much a change in each weight affects the total error 2. Subtract this value from the current weight (multiplied by learning rate) Use a similar process as the output layer, but slightly different to account for the fact that the output of each hidden layer neuron contributes to the output (and therefore error) of multiple output neurons.
- 41. Finally ● The learning rate is how quickly a network abandons old beliefs for new ones. ● High learning rate can lead to over-corrections, and in fact increase the loss instead of decreasing it ● Usually, one can start with a large learning rate, and gradually decrease the learning rate as the training progresses. ● Find a learning rate that is low enough that the network converges to something useful, but high enough that you don't have to spend years training it. 10kiterations
- 42. Debugging Neural Networks 1. Get more training data 2. Play around with regularization techniques 3. Modify network design/capability 4. Train longer 5. Use different types of neurons/networks
- 43. Specialised Neural Networks Convolutional networks used to capture shapes. Used for Computer Vision
- 44. Specialised Neural Networks Recurrent neural network Capture sequential information Used for language modelling
- 45. Learning Takeaway Feature Engineering is implicit Capability Engineering is explicit Humans have been doing this, identify the right animal for a task (dogs, parrots)
- 46. Practical Wisdom ● Machine learning only works if the problem is actually solvable with the data that you have. ● Large amounts of data capture learning that can beat a committee of humans, make sure you supply all nuances of data (shape, color, size) ● Learning is best when training data capture matches serving context. Use data transformation techniques to improve performance (e.g. noise augmentation)
- 48. Thank You!