Artificial Intelligence per Kilowatt-Hour - Max Welling, University of Amsterdam

Overview

Explore the future of energy-efficient artificial intelligence in this thought-provoking lecture by Professor Max Welling from the University of Amsterdam. Delve into the challenges of modern deep learning architectures, which require extensive computational resources and data for training and execution. Learn about innovative approaches to maximize AI capabilities per kilowatt-hour, including neural network compression techniques, low-precision computing, and spiking neural networks. Gain insights into the potential shift of AI computation to edge devices and understand the importance of energy efficiency in the next phase of AI development. Discover how these advancements could shape the future of machine learning across various applications, from medical imaging to reinforcement learning.

Syllabus

Intro
Overview
How to Train a Computer
Deep Convolutional Networks
Resnets
Example: Dermatology
Example: Pathology
Example: Retinopathy
Transfer Learning
Recurrent Neural Networks
Discriminative or Generative?
Latent Variable Models
Amortized Variational Inference
Discriminative vs. Generative Learning
Unsupervised Deep Learning
Image Analogies
Deep Reinforcement Learning
Al is not for free...
Explosive Growth Deep Neural Networks
Al is moving to the Cloud
Solutions
Full Bayesian Compression
Low Bit-Width DL
Spiking Hardware
Spiking Neural Networks
Conclusions