Recommended
Recommended
More Related Content
What's hot
What's hot (20)
Similar to PPT3: Main algorithms and techniques required for implementing Machine Learning & Deep Learning
Similar to PPT3: Main algorithms and techniques required for implementing Machine Learning & Deep Learning (20)
Recently uploaded
Recently uploaded (20)
PPT3: Main algorithms and techniques required for implementing Machine Learning & Deep Learning
- 2. Table of Content ● Programming vs Learning ● Stages of ML Process ● ML Architecture –Data Acquisition –Data Processing –Data Modeling –Execution –Deployment ● EndtoEnd ML & Analytics Architecture ● Skills needed for ML ● ML Practice ● Techniques of ML –Prediction Algorithm ● Description of ML Algorithms with Problem solving Approaches –Supervised –Unsupervised –Ensemble ● ANN Algorithms ● DL algorithms –Summary of Different DL Architecture with pros and cons
- 16. Linear Regression (Lots of numerical data) Used for applications such as Sales Forecasting, Risk Assessment analysis It is used to showcase the relationship between dependent and independent variables K Nearest Neighbors (Group membership based on proximity) Used in industrial applications Used in handwriting detection Used in image/video recognition tasks. It is a lazy algorithm that takes a nonparametric approach to predictive analysis. Logistic Regression (Target variable is categorical) To identifying risk factors for diseases & planning preventive measures Classifying words as nouns, pronouns, & verbs ● Weather forecasting ● It is used for predictive analysis.
- 17. Decision Tree (If/then/else. Noncontiguous data. Can also be regression) Used in pattern recognition Used in option pricing in finances & identifying disease & risk trends. Robust to errors Can handle missing values nicely Can handle both categorical and numerical variables. Random Forest (Find best split randomly.Can also be regression) Used in industrial applications Used for both classification and regression analysis tasks. Creates a forest with a no. of trees & makes them random. Runs efficiently on large databases. Has high classification accuracy. Support Vector Machine (Maximum margin classifier.Fundamental Data Science algorithm) Used in business applications – Such as comparing the relative performance of stocks over a period of time. Classifying data sets into different classes through a hyperplane. Marginalizes the classes & maximizes the distances between them. It require more accuracy & efficiency of data.
- 18. Principal Component Analysis (PCA) (Distil feature space into components that describe greatest variance ) Used in gene expression analysis, stock market predictions Used in pattern classification tasks that ignore class labels. It is dimensionality reduction algorithm. Used for speeding up. Used for making compelling visualizations of complex datasets. It identifies patterns in data & aims to make correlations of variables in them. KMeans Clustering (Similar datum into groups based on centroids) Used in grouping images into different categories, detecting different activity types in motion sensors For monitoring whether tracked data points changes between different groups over time. It works by categorizing unstructured data into a no. of different groups. Each dataset contains a collection of features Algorithm classifies unstructured data & categorizes them based on specific features. Naive Bayes Classifier (Updating knowledge step by step with new info ) Based on Bayes Theorem of probability. Used in document classification,spam filters, sentiment analysis etc. These algorithm for ranking pages, indexing relevacy scores and classifying data categorically.
- 20. NAME DESCRIPTION ADVANTAGES DISADVANTAGES LINEAR REGRESSION The “best fit” line through all data points. Predictions are numerical. Easy to understand – we clearly see what the biggest drivers of the model are. ● Sometimes too simple to capture complex relationships between variables. ● Tendency for the model to “overfit”. LOGISTIC REGRESSION The adaptation of linear regression to problems of classification (e.g., yes/no questions, groups, etc.) Also easy to understand. ● Sometimes too simple to capture complex relationships between variables. ● Tendency for the model to “overfit”. DECISION TREE A graph that uses branching method to match all possible outcomes of a decision. Easy to understand and implement. ● Not often used on its own for prediction because it’s also often too simple and not powerful enough for complex data.
- 21. NAME DESCRIPTION ADVANTAGES DISADVANTAGES RANDOM FOREST Takes the average of many decision trees, each of which is made with a sample of the data. Each tree is weaker than a full decision tree, but by combining them we get better overall performance. A sort of “wisdom of the crowd”. Tends to result in very high quality models. Fast to train. ● Can be slow to output predictions relative to other algorithms. ● Not easy to understand predictions. GRADIENT BOOSTING Uses even weaker decision trees, that are increasingly focused on “hard” examples. High – performing. ● A small change in the feature set or training set can create radical changes in the model. ● Not easy to understand predictions. NEURAL NETWORKS Mimics the behavior of the brain. Neural networks ae interconnected neurons that pass messages to each other. Deep learning uses several layers of neural networks put one after the other. Can handel extremely complex tasks – no other algorithm comes close in image recognition. ● Very, very slow totrain, because they have so many layers. Require a lot of power. ● Almost impossible to understand predictions.
- 23. Description: Computations are structured in terms of interconnected groups, much like the neurons in a brain. Neural networks are used to model complex relationships between inputs and outputs to find patterns in data or to capture a statistical structure among variables with unknown relationships. They may also be used to discover unknown inputs (unsupervised). Usage Examples in Business: ➢ Predicting financial results ➢ Fraud detection Neural network Supervised type
- 24. Description: Computations are structured in terms of categories outputs or observations based on defined classifications. Classification models are used to predict new outputs based on Classification rules. Regression models are generally used to predict outputs from training data. Usage Examples in Business: ➢ Spam filtering ➢ Fraud detection Classification And/or regression Supervised type
- 25. Description: Computations are particular representations of possible solutions to a decision based on certain conditions. Decision trees are great for building classification models because they can decompose datasets into smaller, more manageable subsets. Usage Examples in Business: ➢ Risk assessment ➢ Threat management systems ➢ Any optimization problem where an exhaustive search is not feasible Decision tree Supervised type
- 26. Description: Computations are structured in terms of groups of input data (clusters) based on how similar they are to one another. Cluster analysis is heavily used to solve exploratory challenges where little is known about the data. Usage Examples in Business: ➢ Financial transactions ➢ Streaming analytics in IoT ➢ Underwriting in insurance Cluster analysis Unsupervised type
- 27. Description: Computations are used to provide a description or label to input data, such as in classification. Each input is evaluated and matched based on a pattern identified. Pattern recognition can be used for supervised learning as well. Usage Examples in Business: ➢ Spam detection ➢ Biometrics ➢ Identify management Pattern recognition Unsupervised type
- 28. Description: Computations are rulebased in order to Determine the relationship between different Types of input or variables and to make Predictions. Usage Examples in Business: ➢ Security & intrusion detection ➢ Bioinformatics ➢ Manufacturing and assembly Association rule learning Unsupervised type
- 29. Ensemble Algorithms Ensemble methods are models composed of multiple weaker models that are independently trained & whose predictions are combined in some way to make the overall prediction. Techniques are: ● Boosting ● Bootstrapped Aggregation (Bagging) ● Stacked Generalization (Stacking) ● Gradient Boosting Machines (GBM) ● Gradient Boosted Regression Trees(GBRT) etc.
- 30. Neural Networks (Complex relationships. Prone to overfitting Basically magic) • DL networks used in handwriting analysis, colorization of black & white images • Computer vision processes • Describing or captioning photos based on visual features ANN • It consist of different layers which analyze data. • Hidden layers detect patterns in data. • Greater the no. of layers, more accurate the outcomes. • Neural networks learn on their own & assign weights to neurons every time their networks process data.
- 31. Artificial Neural Network Algorithms: ANN are models that are inspired by the structure and/or function of biological neural networks. They are a class of pattern matching that are commonly used for regression and classification problems but are really an enormous subfield comprised of hundreds of algorithms and variations for all manner of problem types. Algorithms are: ● Perceptron ● Multilayer Perceptrons (MLP) ● BackPropagation ● Stochastic Gradient Descent ● Hopfield Network ● Radial Basis Function Network (RBFN)
- 32. Perceptron • It is a basic processing unit. • The output of a perceptron is the weighted sum of its inputs and a bias unit, which acts as an intercept Multilayer Perceptron • A multilayer perceptron, besides input & output layers, also has hidden layers. • This stack of layer allow it to learn nonlinear decision boundaries in the data. Stacked Autoencoder • It is multiple layers of autoencoders that are trained in an unsupervised fashion individually. • After this one final softmax layer is trained.
- 33. Deep Learning Algorithms: DL methods are a modern update to ANN that exploit abundant cheap computation. They are concerned with building much larger and more complex neural networks. They are concerned with very large datasets of labelled analog data, such as image, text, audio, and video. Algorithms are: ● Convolutional Neural Network (CNN) ● Recurrent Neural Networks (RNNs) ● Long ShortTerm Memory Networks (LSTMs) ● Stacked AutoEncoders ● Deep Boltzmann Machine (DBM) ● Deep Belief Networks (DBN) ● Generative Adversarial Network (GAN) ● Deep Learning based Transfer Learning
- 34. CNN & RNN • CNN are feedforward Neural networks which take in fixed inputs & give fixed outputs. • For example image feature classification,video processing • RNN use internal memory • RNN are versatile • Use timeseries information for giving outputs. • For example language processing tasks, text & speech analysis Deep Belief Networks • Used in field of Image Recognition, Video Sequence recognition, and Motioncapture data. • It is comprised of multiple layers of graphical models having directed and undirected edges. • DBN does not use any labels. • DBNs are generative models. • A DBN finetunes the entire input in a sequence as the model is trained
- 35. Boltzmann Machine • These are two layer neural networks which make stochastic decisions • It does not discriminate between neurons • It learns the distribution of data using the input & makes inferences on unseen data. • It is a generative model – it does not expect input, it rather creates it. Generative Adversarial Networks • GANs are used for generating new data. • GAN comprises of 2 parts, a discriminator and a generator. • Generator is like a reverse CNN, it takes a small amount of data & up scales it to generate input. • Discriminant takes this input & predicts whether it belongs to the dataset. • GANs have been used to generate paintings.
- 37. Convolutional Neural Network Description ● Inspired by the neuralbiological model of the visual cortex. ● Wellsuited for 2D data, i.e., images, thus 1D temporal data need to be preprocessed to form 2D vectors. ● ReLU after convolutional layers help accelerate the convergence speed. Characteristics Pros: – Few neuron connections required with respect to a typical ANN. – Many variants have been prposed: ADCNN, LiftingNet, and inception net, etc. – The classical CNN exhibits a good denoising capability. Cons: – May require many layers to find an entire hierarchy. – May require a large labeled dataset.
- 38. Deep Autoencoder Description ● Mainly designed for feature extraction or dimension reduction. ● Has the same number of input and output nodes. ● Unsupervised learning method aiming at reconstructing the input vector. Characteristics Pros: – Does not require labeled data. – Many variations have been proposed to make the representation more noiseresilient and robust: stacked denoising AE, deep ensemble AE, and stacked sparse AE. Cons: – Requires a pretraining stage. – Training may suffer from the vanishing of errors.
- 39. Deep Belief Network Description ● Composed of RBMs where each subnetwork’s hidden layer serves as the visible layer for the text. ● Has undirected connections just at the top layers. ● Allow unsupervised and supervised training of the network. Characteristics Pros: – Proposes a layerbylayer greedy learning strategy to initialize the network. – Tractable inferences maximize the likelihood directly. Cons: – Training may be computationally expensive due to the initialization process and sampling stage.
- 40. Recurrent Neural Network Description ● An ANN capable of analyzing 1D sequential or temporal data streams. ● LSTM revibrated the application of RNNs. ● Suitable for applications where the output depends in the previous computations. Characteristics Pros: – Memorizes sequential events. – Models time dependencies. – Capable of receiving inputs of variable lengths. Cons: – Frequent learning issues due to gradient vanishing/exploding.
- 41. Generative Adversarial Network Description ● Mainly designed to generate photographs that look superficially authentic to human observers. ● Implemented by a system of generative and discriminative networks. ● SemiSupervised learning method aiming at reconstructing the input vector. Characteristics Pros: – Requires almost no modifications when transferring to new applications. – Requires no Monte Carlo approximations to train. – Does not introduce deterministic bias. Cons: – GAN training is unstable as it requires finding the Nash equilibrium of a game. – Hard to learn to generate discrete data, such as text.