ASGE at DDW AI in GI Workshop | May 2022-Evaluating AI Devices: What to Look For?

Pdf Summary

This document discusses the concept of scale in computer science and machine learning. It begins by stating that scale has changed computer science, using the example of the Apollo 11 Guidance Computer (AGC) that was able to fly, land, and return from the moon, compared to the Google Pixel phone that only charges a phone. It then mentions the scale in terms of hardware components, such as clock speed, RAM, and program size, comparing the AGC's specifications to modern devices.<br /><br />Next, the document discusses how scale is changing machine learning. It mentions the Transformer and Perceptron models and their initial hardware implementation on a Mark I device. It also highlights the progression of accuracy in machine learning, stating that in 2015, the error rate was 4.5%, compared to human error rates of 16% in the same task.<br /><br />The document then mentions the applications of AI, such as real-time audio speech translation for multiple languages and AlphaGo, an AI that beat the world champion in the game of Go. It also introduces AlphaZero, an algorithm that can be applied to various domains.<br /><br />The concept of AutoML is mentioned as a way to build and deploy custom machine learning models with minimal effort and expertise required. It also mentions the exponential growth in AI computation, illustrated by a report from OpenAI, and the progress made in TPUs (Tensor Processing Units) with increasing petaflops of computing power.<br /><br />The document concludes with a thank you note and contact information for Massimo Mascaro.

Asset Subtitle

Massimo Mascaro, PhD

Keywords

scale

computer science

machine learning

AGC

hardware components

accuracy

AI applications

AlphaGo

AutoML

TPUs