The AI Trinity - Data + Algorithms + Infrastructure

Explore the fundamental components driving artificial intelligence in this 53-minute lecture from the Simons Institute Open Lecture Series. Delve into the AI Trinity of data, algorithms, and infrastructure as presented by Anima Anandkumar from the California Institute of Technology. Learn about named entity recognition, active learning with partial feedback, and crowdsourcing techniques for data annotation. Discover the power of data augmentation through generative modeling and symbolic expressions. Gain insights into neural rendering models and tree LSTM architectures. Understand the importance of integrating these three pillars for advancing AI technology, with examples from industry leaders at NVIDIA.

Intro
TRINITY FUELING ARTIFICIAL INTELLIGENCE
TASK: NAMED ENTITY RECOGNITION
RESULTS NER task on largest open benchmark (Onto-notes)
ACTIVE LEARNING WITH PARTIAL FEEDBACK
RESULTS ON TINY IMAGENET (100K SAMPLES) Accuracy vs. Mof Questions
TWO TAKE-AWAYS
CROWDSOURCING: AGGREGATION OF CROWD ANNOTATIONS
PROPOSED CROWDSOURCING ALGORITHM
LABELING ONCE IS OPTIMAL: BOTH IN THEORY AND PRACTICE
DATA AUGMENTATION 1: GENERATIVE MODELING
PREDICTIVE VS GENERATIVE MODELS
STATISTICAL GUARANTEES FOR THE NRM
NEURAL RENDERING MODEL (NRM)
NEURAL DEEP RENDERING MODEL (NRM)
DATA AUGMENTATION 2: SYMBOLIC EXPRESSIONS
ARCHITECTURE: TREE LSTM
SOME RESEARCH LEADERS AT NVIDIA
CONCLUSION Al needs integration of data, algorithms and infrastructure