Generative AI for Beginners AI, ML & LLM Fundamentals

Discover how machine learning builds algorithms from historical data by recognizing patterns and predicting outputs, contrasting legacy programming with data-driven input-output learning.

Identify and compare supervised and unsupervised machine learning, emphasizing labeled versus unlabeled data, and explore regression, binary and multi-class classification, plus clustering algorithms.

Choose the right algorithm for your data to build a model, and learn linear and logistic regression, naive Bayes, decision trees, support vector machine, artificial neural network, and clustering.

Discover how a model is a mathematical representation of data patterns built by an algorithm, using feature sets for prediction, classification, and grouping through training and evaluation.

Learn how machine learning models are trained using features and labels, creating a model through supervised learning, and evaluating via training and testing splits with randomization to ensure prediction quality.

Evaluate model performance using metrics like accuracy, precision, recall, and f1, confusion matrix, and error measures such as mae, mse, and rmse to guide deployment decisions.

Define the problem, collect diverse data from multiple sources, clean and engineer features, select algorithms, train and evaluate models, deploy the best one, and monitor performance over time.

Explore how neural networks mimic the brain by processing through layered nodes with weights, biases, and activation functions to decide, and when to use them over machine learning.

Learn about the perceptron, a single neuron from 1958. Understand how weighted inputs, a bias, and a step activation function implement binary classification.

Explore how artificial neural networks use features, weights, bias, and activation functions across layers to perform forward and backward propagation, train with gradient descent, and produce predictions with confidence score.

Explore deep neural networks with multi-layer architectures and automatic feature extraction from raw data, powered by GPUs, TPUs, public data sources, and Python frameworks.

Explore fully connected dense neural networks, where every neuron connects to the next layer, revealing heavy compute and 625-input image examples that make them impractical for images.

Convolutional neural networks extract image features with 3x3 kernels, ReLU, and pooling to reduce inputs for a classifier, enabling image classification, object detection, and OCR.

Explore recurrent neural networks and their memory-based looping for sequential data such as language modeling and time series, addressing vanishing gradients and the shift to transformers.

Autoencoders use unsupervised learning to compress data into a latent space via an encoder and decoder, enabling reconstruction and noise reduction in image and video processing.

Explore variational autoencoders, an enhanced encoder that learns the input distribution through mean and variance in the latent space to generate new content and variations.

Generative AI uses deep learning to create new content: text, images, video, and audio, beyond prediction. It enables automated coding, writing, and design with ethical, responsible use and evolving prompts.

Discover the landscape of generative AI models, including LLMs, chatbots, open-source options, and multimodal generation tools for text, image, audio, and video.

Explore how generative ai drives productivity across healthcare, finance, retail, education, and entertainment by automating tasks, generating synthetic data, and enabling prompt-based content creation.

The lecture explains how a generative adversarial network uses a generator and a discriminator to create fake but realistic data from a latent space.

This lecture explains diffusion models that add noise to images and reverse denoise to recreate high-quality images, highlighting text-to-image, video, and audio with transformers.

Explore how transformers enable long-range context and parallel processing, overcoming RNN and LSTM limitations. Learn the encoder-decoder architecture, tokenization, embedding, positional encoding, attention mechanisms, and examples like BIRD and GPTs.

Learn how language models construct a dictionary of tokens from words, assign ids to each token, and grow vocabulary as training data increases.

Learn how embeddings convert word tokens into multi-dimensional vectors that reveal semantic relationships, with cosine similarity measuring closeness between words like cat and kitten and distant words like skateboard.

Learn how transformers use token and positional embeddings with sine and cosine encoding, then apply encoder self-attention and multi self-attention to guide the decoder in predicting the next word.

Explore large language models, generative ai based on transformer architecture, pre-trained on vast text data to generate human-like text, and fine-tuned for summarization, q&a, and coding.

Learn how small language models, transformer-based on-device AI, offer domain-focused, cost-efficient, privacy-preserving alternatives to large models, with millions to billions of parameters and fast GPU-free inference.

Explore open source large language models, their 20b–120b parameters, and how weights, embeddings, activation functions, and transformer architecture enable chain-of-thought and function calling without paying token costs.

Learn how foundation models are generic and how fine tuning with private, proprietary data creates organization-specific custom models.

Fine tuning adapts a pre-trained model to a new domain using small task-specific data, delivering faster training and lower costs than training from scratch. Multishot prompt engineering enables this adaptation.

Trace OpenAI’s evolution from the 2015 founding by Elon Musk and Sam Altman to a capped-profit model, culminating in ChatGPT’s 2022 breakthrough and early OpenAI API developers using GPT-2.

Explore OpenAI models like GPT and Codex, and consider the OpenAI API for building applications, while noting Anthropic's developer-focused tools and the role of other models beyond text-to-speech and Whisper.

Discover Microsoft Copilot across Bing, Word, Excel, PowerPoint, Windows, and GitHub Copilot, plus Microsoft Foundry for building AI apps, including agentic applications, with non-developers able to fine-tune and deploy models.

ai agents autonomously perceive, analyze data, and act using tools via a large language model. they work in agentic workflows, planning tasks, iterating with feedback, and coordinating multi-agent automation.

Examine ethical concerns and nature of generative ai, including deepfakes, identity scenarios, copyright, and bias. Assess accountability and privacy as ai hallucinations and black box decisions challenge originality and safety.

Apply Microsoft’s six responsible AI principles—fairness, reliability, inclusiveness, transparency, accountability, and liability—while prioritizing privacy, security, HIPAA, GDPR, and PII compliance, thorough testing, resilience, and enhancing human capabilities.