Tensorflow Tutorial: Hands-on AI development with Tensorflow

Explore how decision trees create a machine learning algorithm for classification and regression, using entropy, information gain, and conditional entropy to build predictive trees.

Implement a decision tree on the iris data set using a library, split data into training and test sets, train the classifier, and predict and compare results with actual labels.

Learn to use linear regression for predicting house prices from features, training models with gradient descent to minimize the loss and find the best fit line.

Implement linear regression from scratch in Python using a population vs profit dataset. Learn preprocessing, cost computation, gradient descent optimization, visualization, and predicting profits with a best-fit line.

Logistic regression, a supervised classifier, uses a sigmoid on Z = tita^T x to create nonlinear decision boundaries and yield probabilities for binary classification.

Implement logistic regression using numpy and matplotlib to predict university admission from two exam scores, including data preprocessing, sigmoid, cost and gradient, normalization, gradient descent, and visualization.

Explore overfitting and regularization in machine learning, and learn how lambda, L1 and L2 regularize the loss to achieve a just-right model.

Explore model evaluation theory through the confusion matrix for binary classification, and learn how true positives, false positives, true negatives, and false negatives inform accuracy, precision, and recall.

Implement model evaluation using precision, recall, and confusion matrix with the metrics library. Print the confusion matrix and the classification report from training data to assess a decision tree model.

Introduce neural networks, showing how inputs combine with weights, produce z, and pass through activation functions like sigmoid to create nonlinear decision boundaries; compare to logistic regression.

Explore how neural networks implement logic gates and forward propagation, building an and gate and an or gate, and highlight XOR limitations and feature engineering.

Explore complex non-linear decision boundaries formed by feature engineering and neural networks. See how neural models automate feature selection, using XOR demonstrations and multi-output architectures for classification.

Discover how back propagation computes gradients for neural networks, enabling gradient descent to optimize weights and minimize loss through forward and backward passes.

Explore gradient descent types batch, stochastic, and mini-batch, and learn how softmax converts neural network outputs into normalized probabilities for multi-class classification.

Build a neural network for digit classification with TensorFlow on a 28x28 grayscale dataset (60,000 training, 10,000 test images), using flattening, dense layers, and a softmax output.

Explore convolutional neural networks for image and video processing, covering classification, detection with bounding boxes, localization, and challenges in shape and appearance variation, illumination, clutter, and occlusion.

Learn how convolution in CNN detects features by sliding a feature matrix over an input image, summing products, using zero padding, and stacking feature maps into depth.

Explore convolution in CNNs, covering filter depth alignment with input, feature maps, output channels equal to number of filters, stride and padding, and max/average pooling to reduce dimensions.

Explore the layers of cnn, including convolution, relu activation, and pooling, and learn to stack them with fully connected and softmax output to classify images.

this lecture demonstrates digit classification using a convolutional neural network on the mnist dataset, achieving 97% training and 97.9% validation accuracy with two conv layers, max pooling, and dense layers.

Explore famous cnn architectures such as LeNet-5, VGG-16, ResNet-34, and GoogleNet, and learn how inception modules and skip connections boost performance and efficiency.

Compare supervised and unsupervised learning, and introduce k-means as an unsupervised clustering method. Initialize random centroids, assign points to the nearest centroid, update centroids as means, and iterate until convergence.

Explore the k-means algorithm in unsupervised learning: initialize k centroids, assign points to the nearest centroid to form clusters, update centroids by averaging, and repeat until convergence.

Explore centroid initialization in k-means, showing how random or poor initial centroids slow convergence and worsen clustering, then evaluate multiple initializations by cost to select the minimum-cost option.

Apply k-means++ to smartly initialize centroids by selecting far-apart data points, ensuring each centroid belongs to a different cluster and speeding convergence.

Explore how to determine the optimal number of clusters using the elbow method, analyze error versus cluster count, and lay the groundwork for implementing k means.

Implement the k-means algorithm on a custom dataset X, specifying the number of clusters and random_state for centroid initialization, then obtain labels, predict points, and view cluster centers.

Learn principal component analysis for data reduction and dimensionality reduction, preserving maximum information while rotating axes to reveal the principal components for applications like image compression and facial recognition.

Learn how to perform facial recognition by applying principal component analysis to reduce image dimensions, then classify with a multilayer perceptron, blending unsupervised and supervised learning.

Train a convolutional neural network in TensorFlow to recognize fashion items from Fashion-MNIST, using two convolutional layers, pooling, flattening, and three dense layers, achieving about 85 percent validation accuracy.

Train a cnn on cifar-10 to classify 10 color classes and observe limitations of standard techniques; plan to boost accuracy with architecture tweaks and transfer learning in the next video.

Classify cats vs dogs using transfer learning with a pre-trained MobileNet v2, and combine it with custom layers. Prepare, reshape, normalize, and train with a frozen base.

Apply action recognition on YouTube videos by sampling frames, extracting features with a pretrained Inception V3, and classifying 101 actions using an LSTM model with transfer learning.