CS 585: Lecture 8, Machine Learning

Data-driven approach to AI

Machine learning is a data-driven subset of AI built on statistics, linear algebra, and calculus, using symbolic, reinforcement, and neural approaches to learn from data (supervised/unsupervised/semi-supervised) to optimize performance and business outcomes.

Supervised, unsupervised, semi-supervised learning

Supervised learning maps labeled inputs to outputs, unsupervised learning discovers hidden structure/manifolds in unlabeled data, and semi-supervised learning combines a small labeled set with lots of unlabeled data so structure plus labels guide the model.

Reinforcement learning (RL)

RL models an agent interacting with an environment to learn a policy that maximizes long-term reward, often via MDPs/POMDPs and hierarchical decompositions, and is used for robotics, navigation, and sequential decisions.

Data-driven ML & AI history pointers

Modern deep learning systems train neural networks on massive datasets to find patterns and then generate or extrapolate new data, underpinning vision, self-driving, chatbots, and GenAI, as covered in AI history timelines and documentaries.

Connectionist / neuro / data-driven AI workflow

The typical neural workflow is: collect and (optionally) label data, iteratively train a network by adjusting parameters to reduce loss, then deploy it so it can predict on new inputs, loosely inspired by how biological neurons learn.

Artificial neural networks (ANNs)

ANNs are networks of interconnected, weighted nonlinear units that are trained on labeled examples to learn patterns and then use those patterns to make predictions on new data.

Single neuron – computation

A single neuron computes a weighted sum of its inputs plus a bias and passes it through a nonlinear activation (e.g., sigmoid), producing an output used by downstream neurons or as a prediction.

NN applications & brain analogy

Neural nets can learn patterns in almost any structured data (classification, anomaly detection, forecasting, etc.) and are loosely analogous to the brain’s huge network of neurons and adaptive synaptic weights.

Artificial neurons & activation functions

Different activation functions (linear, threshold, ReLU, sigmoid, etc.) map net input to output and provide the nonlinearity and gradients needed for neural nets to learn complex patterns via backprop.

NN theory & layered architectures

Neural nets can be seen as directed graphs or large dynamic systems where layers of units transform inputs into higher-level features and learning means adjusting parameters in a model y = f(x; W) to minimize a loss.

Backpropagation learning

Backprop iteratively adjusts weights by propagating gradients from the output layer backwards, using hyperparameters like learning rate and momentum to reduce loss and assign “credit/blame” to individual units.

Hierarchical features, universal approximation & simple examples

Multi-layer nets learn hierarchical features (e.g., pixels → parts → faces) and, with enough neurons and the right nonlinearities, can approximate any continuous function, as seen in simple classifiers and regression examples.

XOR, Softmax & why deep NNs matter

Solving non-linearly separable problems like XOR requires multi-layer networks, Softmax extends sigmoid to multi-class output, and deep NNs took off thanks to big data, better algorithms, and powerful hardware.

Neuron summary & implementation

A simulated neuron corresponds to biological dendrites, soma, and axon: it sums weighted inputs, passes them through an activation, and emits an output signal in code.

Sigmoid importance & ML pipeline

Nonlinearities like sigmoid keep networks from collapsing into a single linear mapping, and a typical ML pipeline goes from understanding the business problem through data prep and training to deployment and ongoing monitoring.

Supervised ML as giant nonlinear function & deep learning

Supervised ML can be viewed as learning a big nonlinear function y = F(x) by tuning architecture and hyperparameters so predictions match targets, with deep learning using many layered nonlinear units enabled by modern data and compute.

Deep learning uses, “deepness” & architectures (incl. RNNs)

Deep nets with many hidden layers power state-of-the-art image, speech, and NLP systems using architectures like CNNs, RNNs/LSTMs, TCNs, HTM, and Transformers, where “depth” is the number of intermediate layers.

Convolution & CNNs

Convolution slides small kernels over inputs to extract features, and CNNs stack many such layers (with ReLU and pooling) to learn complex patterns in images, speech, and other signals, enabling systems like YOLO for real-time detection.

GANs, EBMs, VAEs & encoder–decoder pairs

Generative models like GANs (adversarial generator–discriminator), EBMs, and encoder–decoder VAEs learn data distributions and latent spaces so they can generate new realistic samples from noise or latent codes.

Applications & demos – sampler

ML is now deployed across thousands of real-world applications—from vision, speech, medicine, and security to games and creative tools—demonstrating mature, production-ready capabilities.

Problems, hype & critical views

Key concerns include biased data, deepfakes, adversarial weakness, lack of explainability, and concentrated power, leading many researchers to critique current ML/AI limits and hype.

Current directions & major players

Cutting-edge work spans transformers, neuromorphic and optical computing, GNNs, contrastive learning, pruning, and explainability, driven by major tech companies and cloud providers worldwide.