Master Deep Learning: A Comprehensive Bootcamp

Explore fundamentals of deep learning, including neurons, activation functions, gradient descent and stochastic gradient descent, and move from artificial neural networks to CNNs and LSTMs, with libraries and frameworks overview.

Understand what deep learning is within machine learning, its data needs for image classification, and recognize Geoffrey Hinton as the godfather of deep learning.

Explore the basic neural network architecture, including input, hidden, and output layers, and learn how neurons combine inputs with weights to produce y hat through activation functions.

Activate neurons by applying activation functions to the weighted input sum plus bias, deciding whether to fire. Learn about step, linear, sigmoid, tan, and relu functions.

Explore the step function as a threshold-based activation that activates neurons above a set value, enabling binary classification. It highlights its limits for multi-class problems, using churn and iris examples.

Explore the linear activation function y = m x + c and its gradient; use it in one hidden layer since stacking linear activations yields a single linear function.

Explore the sigmoid activation function, a non-linear 0 to 1 output function with high and vanishing gradients, its use in output layers, and its relation to Relu in deep learning.

Explore the tanh function, its -1 to 1 range, and its relation to a scaled sigmoid where tanh equals two times sigmoid function of two x minus one.

Explore the rectified linear unit (Relu) activation function, its non-linear, sparsity-friendly behavior, and its approach to mitigating vanishing gradients in deep networks.

Learn how forward pass and backward propagation train neural networks by updating weights to minimize the cost function across batches. Explore input features, batch size, learning rate, and gradient descent.

Explore gradient descent and stochastic gradient descent, including forward and backward passes, weight updates, and how cost functions guide learning across batches and epochs.

Explore the basics of artificial neural networks, detailing input, hidden, and output layers, feature-driven neuron counts, and a churn analysis workflow comparing models with decision trees and random forests.

Demonstrate churn modeling with practical ML: compare decision tree, random forest, and a keras neural network. Include one-hot encoding, feature scaling, train-test split, and hyperparameter tuning considerations.

Explore hyperparameter optimization techniques, including manual search, randomized search CV, and grid search, to identify optimal epochs, optimizers, and other parameters via cross-validation in a Keras classifier workflow.

Explore deep learning basics, neural networks, activation functions, gradient descent, and forward and backward propagation, plus convolutional neural networks, their 1d, 2d, and 3d variants, and an end-to-end project.

Explore what a convolutional neural network is, why it handles images, voice, and videos, focusing on 2D CNNs and the steps: convolution, relu, pooling, flattening, and a fully connected layer.

Explore how cnn relies on convolution as the core dimensionality reduction step that preserves features, then relu, pooling, flattening, and the fully connected layer for ten-class classification.

Explore how convolutional neural networks detect features with 3x3 to 7x7 feature detectors, apply relu, pooling, flattening, and fully connected layers, and grasp strides and starting with 32 filters.

Learn how image augmentation balances imbalanced datasets by generating new data through upsampling, resampling, and image transformations such as flipping, rotation, noise addition, and zoom.

Explore the core concepts of gradient descent in neural networks and compare epoch, batch size, and iterations. Learn how forward and backward passes update weights to optimize learning.

Learn to build a basic CNN on Fashion-MNIST using Keras and TensorFlow, with Conv2D, max pooling, flatten, and dense layers to classify 10 grayscale 28x28 images.

Understand how a CNN model summary lists layers and trainable parameters as a 28 by 28 input goes through two 3 by 3 convs, pooling, and flattening.

Demonstrates building a binary x-ray classifier for pneumonia versus normal cases using TensorFlow and Keras, with grayscale 100x100 images and a CNN trained and evaluated via provided data.

Explore how recurrent neural networks capture time-dependent patterns by updating hidden states with previous outputs, enabling sequence modeling for time series, text prediction, and stock data.

Explore the four model types—1 to 1, 1 to many, many to one, and many to many—with examples like image classification, image captioning, sentiment analysis, and translation.

Explore vanishing and exploding gradient problems in neural networks, including how sigmoid derivatives affect backpropagation, and learn fixes like Relu, RMSprop, and LSTMs.

Explore lstm, designed to overcome vanishing and exploding gradients, featuring input, forget, and output gates and an internal state that enables short term and long term learning.

Learn how to build an LSTM-based stock price predictor using Keras, with 60-time-step sequences, minmax scaling, and a multi-layer RNN including dropout, trained with RMSprop and mean squared error.

Learn how pre-trained models, trained on large data like ImageNet, can be fine-tuned or used as feature extractors to accelerate deep learning tasks with limited data or resources.

Explore practical x-ray image classification by building CNNs from scratch and with pre-trained vgg16 base models, fine-tuning with custom layers, and comparing training, validation, and test accuracies.

Vgg16 is a pre-trained convolutional neural network for image classification across 1000 classes, featuring 13 conv layers, 5 pooling layers, 3 dense layers, 3x3 filters, stride 1, and 224x224 input.

Explore depthwise and pointwise convolutions in MobileNet, a lightweight Google open-source cnn family, highlighting speed, power efficiency, and architecture across v1–v3 for image classification.

Explore transfer learning by leveraging pre-trained models to fine-tune for your tasks, boosting accuracy and saving training time, while acknowledging bias risks and when to apply.

Apply final project techniques by building and deploying a cnn-based pneumonia detector using pre-trained models, Streamlit web app, and end-to-end workflow from training to deployment.