Deep Learning | Tensor Flow | RBM | Auto Encoders | GAN

Explore tensors, from scalars to vectors and matrices, and learn how TensorFlow graphs capture data flows, operation order, and directed acyclic graphs.

Learn to define tensors, perform arithmetic and reduction operations, reshape, index, and broadcast across two-by-three tensors; explore matrix multiplication, transpose, and eigenvectors and eigenvalues.

Demystify neural networks by comparing factors and weights to shirt buying, then reveal how inputs, weights, activation, and hidden layers predict outcomes through training, loss, backpropagation, and gradient descent.

Investigate activation functions and their role in managing non linearity, exploring sigmoid, ReLU, leaky ReLU, tanh, softmax, and ELU in neural networks.

Explore convex optimization and gradient descent to locate the global minimum of a convex function, using derivatives, loss, and cost functions within deep learning and machine learning.

Develop a churn prediction model with an artificial neural network, evaluate accuracy across activation functions using TensorFlow, and use binary cross entropy, confusion matrix, encoding, and scaling.

Demonstrate how convolution in cnn uses kernels to extract edges and local features into feature maps, then apply activation, pooling, padding, and a fully connected layer with softmax for classification.

Train a CNN model to identify ten image categories from a manifest dataset of grayscale 28 by 28 images using TensorFlow, NumPy, Pandas, and Matplotlib for visualization.

Learn how recurrent neural networks create memory and persistence by looping information through a hidden state. Apply this to time series forecasting and predicting the next word in a sequence.

Learn how an RNN updates its hidden state, or memory, in loops using current inputs, weights, and biases to generate outputs via softmax.

LSTM extends RNNs to capture long-term context with memory cells and a cell state, using the input, forget, and the output gates to address vanishing gradients for better predictions.

Build a spam detection model using rnn and lstm with TensorFlow, Keras, embedding, and dropout. Preprocess text, tokenize, and compare rnn and lstm accuracy (about 98% vs 97%).

Contrast ann, cnn, and rnn architectures by detailing their input, hidden, and output layers. Explain cnn's image processing with pooling and rnn's memory cells for sequence context; note vanishing gradient.

Explore generative adversarial networks, where a generator and discriminator compete in a minimax framework to create synthetic data and enable unsupervised learning and data augmentation.

Learn how Boltzmann machines model visible and hidden node relations with generative data sampling, and how restricted Boltzmann machines and deep belief networks stack to form deep representations.

Auto encoders are unsupervised neural networks that encode input to a latent space and decode it back, minimizing reconstruction error for tasks like image and audio compression and anomaly detection.

Develop a fraud detection model for the insurance industry using three unsupervised models—auto encoder, gan, and deep belief network—on 37 features with preprocessing and anomaly thresholds.