Machine Learning with TensorFlow on Google Cloud

Explore how neural networks learn to recognize patterns, comparing brain computation with computers, and train models on datasets like fashion mnist to classify grayscale images into ten categories.

Build a logistic regression model in TensorFlow with a single sigmoid neuron, using Google Colab, getdummies preprocessing, standard scaler, and a train-test split, then evaluate churn at 0.5.

Stack perceptrons in parallel to produce multiple outputs from the same inputs, such as locating x and y coordinates in an image, and use sequential stacking for non linear classifications.

Explore how gradient descent iteratively finds a function’s minimum by moving along the slope from a random start, selecting the steepest descent until no further decrease.

Explore how gradient descent minimizes cross-entropy loss in neural networks for classification by updating weights and biases through backpropagation, with learning rate controlling step size.

Learn why activation functions bound outputs and introduced non-linearity in neural networks. Explore common types—step, sigmoid, tanh, and relu—and how they apply to hidden vs output layers.

Build a neural network from 28 by 28 inputs, flatten to 784, add two relu dense layers (300 and 100), and a softmax output for ten classes using Keras sequential.

Learn how to save and load Keras models with checkpoints in Google Colab and Drive, using callbacks like save best only and early stopping to optimize training.

Demonstrate how a convolutional layer processes an input image with a receptive field window, a five cross five window, to extract lower-level features and build higher-level representations.

Learn stride in CNNs, shifting the convolution window to shape receptive field overlap, with stride 2 and 4 illustrating how stride affects overlap and the upper layer size.

Explore how stride and padding affect coverage in CNNs, comparing valid padding that ignores border pixels with same padding that adds blank pixels to preserve the receptive field.

Explore how convolutional neural networks process images with channels, distinguishing grayscale single-channel inputs from color images with red, green, and blue channels, each pixel valued 0–255.

Learn how pooling layers in CNNs reduce computation, memory usage, and parameters by aggregating inputs with max or average pooling, with no weights to train.

Build a cnn model for fashion mnist with ten categories by preprocessing: reshape 28x28 grayscale images to 4d, normalize by 255, and split 55k train 5k validation before training.

Observe the model's overfitting: training accuracy at 87–88% outpaces validation around 72–74%, prompting data augmentation with zoom, shear, and rotation before retraining and saving the CNN model.