Machine Learning for Interviews & Research and DL basics

Explore supervised machine learning, distinguishing regression and classification with labeled training data, mapping inputs to outputs, and using models such as regression, decision trees, random forest, SVM, and neural networks.

Explore regression in supervised learning, where the target is a continuous value like car or home price, and compare linear and non-linear models to understand training, testing errors, and overfitting.

Learn how classification uses discrete targets with fixed categories, illustrated by credit scoring with income and savings. Explore applications across structured and unstructured data.

Learn linear regression basics, including univariate regression with bias and weights and its matrix form, and the mean squared error, gradient descent, and the pulse objective for energy forecasts.

Apply gradient descent to univariate linear regression, a convex optimization, updating theta0 and theta1 to minimize the cost. Use epochs and learning rate alpha toward a local optimum.

Master gradient descent tips, including feature scaling and mean normalization (and min-max normalization) to speed convergence in weather forecasting, and review batch, mini-batch, and stochastic variants with learning-rate tuning.

Learn the normal equations as an exact analytical method for linear regression, deriving weights with the equation (X^T X)^{-1} X^T y, and compare to gradient descent's iterative approach.

Learn locally weighted linear regression, a non-parametric method that computes query-point weights from nearby training data using a diagonal weighting matrix and Gaussian bandwidth, with no training phase.

ridge regression adds a penalty term to linear regression, proportional to the squared weights, to reduce overfitting and shrink weights without eliminating them.

Lasso regression adds a linear regularization term to linear regression to reduce overfitting and training error. Unlike standard regression, lasso can set some weights to zero, enabling feature selection.

Explore and train key classification models in sklearn, including logistic regression, KNN, SVM, neural networks, decision trees, random forests, gradient boosting, and lda, using import, fit, predict, and evaluation metrics.

Explore logistic regression as a binary classifier with a sigmoid hypothesis and 0.5 threshold. Minimize negative log-likelihood, note non-convex optimization, and apply one-vs-all or New Balance method for multi-class voting.

Explore how to map nonlinear functions with linear models using basis function expansion, including polynomial, radial basis, and sigmoid options, and address overfitting with regularization.

Analyze overfitting and underfitting, and apply regularization to reduce overfitting by penalties on weights, including l2 and l1 methods, with dropout and dimensionality reduction, to improve generalization.

Discover how support vector machines classify and regress data using linear models, kernel tricks for nonlinear problems, and the maximum-margin hyperplane defined by support vectors.

Explore decision trees and boosting techniques. Learn how entropy and information gain guide splits, pruning strategies, and how gradient boosting and AdaBoost improve accuracy.

Explore how neural networks use forward propagation and backward propagation to compute outputs and update weights through gradient descent, stochastic, and minibatch methods, with activation functions like sigmoid and ReLU.

Outline the requirements for deep learning, including three hidden layers, and demonstrate learning rate, activation, and l1 regularization using TensorFlow Playground, with attention to training loss and overfitting.

Learn to build a deep neural network by sizing the input layer to features or image pixels, using ReLU in hidden layers, and constructing a Keras sequential model for regression.

Explains overfitting in deep neural networks caused by high variance, and how dropout and regularization reduce training variance; shows training-time neuron removal and test-time averaging to improve generalization.

Apply batch normalization to hidden-layer inputs using batch mean and standard deviation to improve accuracy, reduce overfitting, and complement dropout, though it is not compulsory for deep neural networks.

Explore how residual connections enable deeper networks to run faster than shallow models, as shown by 152-layer ResNet outperforming the 19-layer VGG, using skip connections to ease gradient flow.

Explore convolutional neural networks, compare them with dense and locally connected networks, and learn about convolution operations, padding, stride, edge detection, and feature extraction for image classification and forecasting.

Explore maximum pooling in cnn models, using a two-by-two filter with stride two to downsample feature maps, compare to average pooling, and understand how pooling reduces parameters and helps prevent overfitting.

Explore recurrent neural networks with feedback and hidden states, enabling time series and text processing, and learn backpropagation through time and the role of LSTM and GRU units.

Explore how LSTM units use forget, input, and output gates to manage long-term dependencies in time series data, with memory cells and a candidate layer guiding sequence outputs.

Discover gru units, a simplified version of lstm with update and reset gates that blend long term and short term memory, addressing vanishing gradients in time series and nlp.

Explore unsupervised learning through clustering, focusing on k means clustering with centroid assignment, mean reassignment, and convergence, and learn how the elbow method determines the optimal k.

Get started with Python for machine learning by using NumPy arrays, indexing, slicing, shape and reshape; load and plot data with Matplotlib and Pandas, and explore case studies.

Case study demonstrates digits classification with Keras, using a logistic regression baseline and a dense hidden layer, softmax activation, categorical cross-entropy, and the Adam optimizer, with data normalization.

Examine a load forecasting case study using smart meter energy data and weather features, detailing time-series preprocessing and a range of models from regression to deep learning.

Learn multivariable linear regression from data loading and exploration to feature selection, model training, and evaluation using mean squared error, mean absolute error, and R-squared.

Explore object oriented programming concepts with a Python example, covering classes, objects, encapsulation, inheritance, polymorphism, and abstraction to organize machine learning pipelines and improve code reuse.