Deep Learning Overview

Machine learning is a sub-field of artificial intelligence that uses data to train predictive models, following a two-phase flow of computationally expensive offline training and fast, cheap inference in production. The overview covers supervised, unsupervised, semi-supervised, self-supervised, and reinforcement learning, as well as regression, classification, ranking, discriminative vs. generative models, transfer learning, fine-tuning, LoRA, and metric learning with siamese networks.

- Basics: stats /andrewt3000/MachineLearning/blob/master/stats.md , linear algebra /andrewt3000/MachineLearning/blob/master/la.md , calculus /andrewt3000/MachineLearning/blob/master/calculus.md data and features /andrewt3000/MachineLearning/blob/master/data.md training neural networks /andrewt3000/MachineLearning/blob/master/neuralNets.md transformers /andrewt3000/MachineLearning/blob/master/transformer.md LLM /andrewt3000/MachineLearning/blob/master/llm.md reinforcement learning /andrewt3000/MachineLearning/blob/master/rl.md - Legacy: cnn /andrewt3000/MachineLearning/blob/master/cnn.md , rnn /andrewt3000/MachineLearning/blob/master/rnn.md - In progress: - vision transformers /andrewt3000/MachineLearning/blob/master/cv.md Machine Learning is a sub-field of artificial intelligence that uses data to train predictive models. Machine learning problems follow a two-phase flow: Training : The model learns from data — adjusting its parameters to minimize error supervised learning or maximize reward RL . This phase is computationally expensive and done offline. Inference : The trained model is deployed and makes predictions on new, unseen inputs. Parameters are frozen — the model applies what it learned. This phase must often be fast and cheap, since it runs in production e.g., serving predictions in real time . Supervised learning - learns from labeled training data.- svm, knn, random forests, gradient boosting machines /andrewt3000/MachineLearning/blob/master/gbm.md , neural networks /andrewt3000/MachineLearning/blob/master/neuralNets.md - svm, knn, random forests, Unsupervised learning - learns from unlabeled training data.- principal component analysis, clustering. - An Reinforcement learning agent interacts with an environment and learns to take action by maximizing a cumulative reward .- Q-Learning, Deep Q-Networks DQN , Proximal Policy Optimization PPO Regression - predicting a continuous value attribute.- Example: predicting house prices Classification - predicting a discrete value.- Classification is further categorized as binary or multi-class classification. - Binary Example: predicting pass or fail, benign or malignant, spam or not spam, hot dog or not hot dog :- - Multi-Class Example: Handwritten Digit Recognition 0 through 9 mnist https://huggingface.co/datasets/ylecun/mnist , Image classification with 1,000 classes ImageNet-1k https://huggingface.co/datasets/ILSVRC/imagenet-1k Ranking - predicting the relative order or preference of a set of items contextually.- Example: search engine results, or movie recommendations Models that predict labels from inputs as in the problems above are called discriminative ; models that learn the data distribution to synthesize new samples diffusion, LLMs are generative . - Transfer learning is storing knowledge gained while solving one problem and applying it to a different but related problem. fine tuning is additional training to a base model for a specific task. LoRA Low-Rank Adaptation is a fine-tuning method that freezes the base model's weights and trains small low-rank matrices that are added to existing layers. - Semi-Supervised learning trains on a mix of mostly unlabeled with a small labeled subset data. - Self-supervised learning is a form of unsupervised learning where training labels are constructed automatically from the data itself.- Autoregressive LLM /andrewt3000/MachineLearning/blob/master/llm.md pretraining next word prediction , and masked image modeling. - Autoregressive - Metric learning trains a model to produce embeddings where distance reflects similarity. A siamese network passes two inputs through identical networks with shared weights and compares the resulting embeddings. Trained with contrastive loss pull matching pairs together, push non-matching pairs apart or triplet loss anchor, positive, negative .- Example: face verification, signature verification