And Then There Are Algorithms
summary
Claims a library of scalable algorithms — k‑means, factorization machines, time‑series forecasting, linear regression, topic modeling, image classification — that scale to any amount of data and run distributed or streaming so training is incremental instead of full retrain. Details include using alternating least squares for factorization, recurrent nets with attention for seq2seq, embeddings for words, and convolution / capsule nets for images. Tooling is SageMaker with TensorFlow MXNet PyTorch containers, model APIs and edge deploy to Raspberry Pi and Nvidia Jetson, tested in production with use cases like musiXmatch recommendations and Vodafone driver monitoring at the edge.
AI-written summary · about
chapters (23)
- 0:00Intro
- 2:45What is an Algorithm?
- 3:48The Master Algorithm
- 4:14The Five Tribes of Machine Learning
- 4:39Linear Learner
- 5:00Bike Sharing Prediction (Regression)
- 6:55Minimizing the Error
- 7:28Objective Function
- 8:57Stochastic Gradient Descent (SGD)
- 12:09Factorization Machines (4)
- 15:16Image Classification
- 15:43Convolutional Neural Networks (CNN)
- 17:29Sequence to Sequence (seq2seg)
- 20:09K-Means Clustering
- 22:14Principal Component Analysis (PCA)
- 23:59Latent Dirichlet Allocation (LDA)
- 25:43Neural Topic Model INT
- 26:14Time Series Forecasting (DeepAR)
- 30:17Model Selection (Hyperparameters)
- 31:35Incremental Training with Streaming
- 32:08Model Selection with Streaming
- 32:28Amazon Sage Maker
- 34:05Machine Learning Inference at the Edge
related talks
19mDeveloping large-scale machine learning algorithms on Amazon SageMaker
53mData Engineering and AI
30mChoose the right machine learning algorithm in Amazon SageMaker
46mMachine Learning with Kubernetes
41mInfrastructure for ML Applications
20mAI and Machine Learning for developers
44mScaling ML from 0 to millions of users
22mReal-world applications of machine learning: Beyond the hype