Yuxi_Liu

No nonsense version of the "racial algorithm bias"

In discussions of algorithm bias, the COMPAS scandal has been too often quoted out of context. This post gives the facts, and the interpretation, as quickly as possible. See this for details. THE FIGHT The COMPAS system is a statistical decision algorithm trained on past statistical data on American convicts. It takes as inputs features about the convict and outputs a "risk score" that indicates how likely the convict would reoffend if released. In 2016, ProPublica organization claimed that COMPAS is clearly unfair for blacks in one way. Northpointe replied that it is approximately fair in another way. ProPublica rebukes with many statistical details that I didn't read. The basic paradox at the heart of the contention is very simple and is not a simple "machines are biased because it learns from history and history is biased". It's just that there are many kinds of fairness, each may sound reasonable, but they are not compatible in realistic circumstances. Northpointe chose one and ProPublica chose another. THE MATH The actual COMPAS gives a risk score from 1-10, but there's no need. Consider the toy example where we have a decider (COMPAS, a jury, or a judge) judging whether a group of convicts would reoffend or not. How well the decider is doing can be measured in at least three ways: * False negative rate = (false negative)/(actual positive) * False positive rate = (false positive)/(actual negative) * Calibration = (true positive)/(test positive) A good decider should have false negative rate close to 0, false positive rate close to 0, and calibration close to 1. Visually, we can draw a "square" with four blocks: * false negative rate = the "height" of the false negative block, * false positive rate = the "height" of the false positive block, * calibration = (true positive block)/(total area of the yellow blocks) Now consider black convicts and white convicts. Now we have two squares. Since they have different reoffend rates for some reason, the

115Jul 13, 2019

Yuxi_Liu

Message

Yuxi Liu is a PhD student in Computer Science at the Berkeley Artificial Intelligence Research Lab, researching on the scaling laws of large neural networks.

Personal website: https://yuxi-liu-wired.github.io/

642

Predicting AGI by the Turing Test

Abstract This essay explains the Direct Approach proposed by [@barnettScalingTransformativeAutoregressive2023].[1] I encourage you to play with the Direct Approach Interactive Model to explore an interactive simulation using the approach. > The Direct Approach framework bounds the compute requirements for transformative AI by extrapolating neural scaling laws. We combine those estimates...

Jan 22, 202421

The Perceptron Controversy

Connectionism died in the 60s from technical limits to scaling, then resurrected in the 80s after backprop allowed scaling. The Minsky–Papert anti-scaling hypothesis explained, psychoanalyzed, and buried. I wrote it as if it's a companion post to Gwern's The Scaling Hypothesis.

Jan 10, 202465

Cybernetic dreams: Beer's pond brain

"Cybernetic dreams" is my mini series on ideas from cybernetic research that has yet to fulfill their promise. I think there are many cool ideas in cybernetics research that has been neglected and I hope that this series brings them more attention. Cybernetics is a somewhat hard to describe style...

Nov 19, 201939

An optimal stopping paradox

Consider an optimal stopping problem: a company at each time step grows by some constant, and has a certain probability of shutting down. You decide when to sell the company. Since the math is cleaner in continuous time, we consider the continuous time. Then the company has a linearly increasing...

Nov 12, 201911

Living the Berkeley idealism

Quick observation, more funny than insightful. Today I was thinking about how to publish my thesis when it's finished, and rethinking again what format to put it in. The standard pdf format seems reasonable, but it is merely made of digital print, with no interactivity. Putting in interactivity requires me...

Sep 4, 201912

Why study perturbative adversarial attacks ?

This post summarizes and comments on Motivating the Rules of the Game for Adversarial Example Research Summary of paper Despite the amount of recent work done, human-inperceptible perturbation adversarial attacks (Example: One Pixel Attack) are not as useful as the researchers may think, for two reasons: 1. They are not...

Aug 29, 201919

How to take smart notes (Ahrens, 2017)

This is my rephrasing of (Ahrens, 2017, How to Take Smart Notes). I added some personal comments. The amazing note-taking method of Luhmann To be more productive, it's necessary to have a good system and workflow. The Getting Things Done system (collect everything that needs to be taken care of...

Jul 23, 201972

Load More (7/16)

LESSWRONG
LW

LESSWRONG
LW

Yuxi_Liu

Yuxi_Liu

Yuxi_Liu

No nonsense version of the "racial algorithm bias"

How to take smart notes (Ahrens, 2017)

The Perceptron Controversy

Let's Read: Superhuman AI for multiplayer poker

Yuxi_Liu

Predicting AGI by the Turing Test

The Perceptron Controversy

Cybernetic dreams: Beer's pond brain

An optimal stopping paradox

Living the Berkeley idealism

Why study perturbative adversarial attacks ?

How to take smart notes (Ahrens, 2017)

No nonsense version of the "racial algorithm bias"

How to take smart notes (Ahrens, 2017)

The Perceptron Controversy

Let's Read: Superhuman AI for multiplayer poker

Predicting AGI by the Turing Test

The Perceptron Controversy

Cybernetic dreams: Beer's pond brain

An optimal stopping paradox

Living the Berkeley idealism

Why study perturbative adversarial attacks ?

How to take smart notes (Ahrens, 2017)