Last Week in AI #95
Making AI more energy efficient, regulating AI hiring tools, and more!
In a recent paper published at NeurIPS, the largest annual AI conference, researchers from IBM unveiled an algorithm that can train deep neural networks with only 4-bit precision numbers. For context, neural networks can be trained with numbers of varying precisions, from double (64 bit) to single (32 bit), and more recently half (16 bit). "Precision" roughly corresponds to how many possible numbers each variable in the neural network can represent, and consequently how much memory and computational resources the neural network needs. With 4 bits, each variable can only take on 2^4 = 16 values, which is tiny compared to 2^16 = 65536 values for half precision training.
The IBM researchers simulated 4-bit precision training on a variety of deep learning tasks, from image recognition to natural language processing. Their algorithm did not perform significantly worse than half-precision training and was 7 times faster and more energy efficient. While 4-bit precision training in the real world will not happen for another few years (this requires 4-bit precision hardware, which is in development), this is still an exciting result with many potential applications. Low-precision training could make deep learning more resource and energy efficient.
It could also make training powerful AI models possible on smartphones and other small devices, which would improve privacy by helping to keep personal data on a local device. And it would make the process more accessible to researchers outside big, resource-rich tech companies.
AI hiring tools have gained popularity in recent years. While they have the potential to make the hiring process more efficient and effective for both employers and job-seekers, there are real concerns over recent "cases of discrimination based on gender, race and disability during candidate sourcing, screening, interviewing and selecting using automated tools." As such, New York City Council introduced a bill that would require employers inform candidates if automated-decision systems were used and also require AI technology vendors to provide bias audits before the products can be sold. If passed, the bill would take effect in 2022.
Julia Stoyanovich, professor of computer science at New York University's Center for Responsible AI, suggests that employers should also allow "the job applicant to understand, and, if necessary, correct and contest the information" given to automated hiring systems. With regard to the bill, she notes how "if left unchecked, automated hiring tools will replicate, amplify and normalize results of historical discrimination in hiring and employment."
Advances & Business
Uber sells self-driving unit Uber ATG in deal that will push Aurora's valuation to $10B - "Aurora Innovation, the autonomous vehicle startup backed by Sequoia Capital and Amazon, has reached an agreement with Uber to buy the ride-hailing firm's self-driving unit in a complex deal that will value the combined company at $10 billion."
Hyundai to acquire Boston Dynamics for nearly $1B - "Hyundai Motor will acquire Boston Dynamics. The acquisition will be finalized at Hyundai�s�December 10 board meeting. News about the deal was first reported by The Korea Economic Daily, which said the deal is for $921 million (1 trillion won)."
Banks look at "explainable" AI systems to boost consumer trust - "Fraud detection is another area in which AI tools are extensively used by banks. This has been the case for many years, but the tools are becoming smarter and more focused, on areas like anti-money laundering, due diligence, and portfolio credit risk optimization."
OcéanIA treats climate change like a machine learning grand challenge - "Now a team of researchers with the recently established OcéanIA is treating the study of the ocean and climate change as a machine learning grand challenge."
Concerns & Hype
Police Drones Are Starting to Think for Themselves - "When the Chula Vista police receive a 911 call, they can dispatch a flying drone with the press of button."
The U.S. Air Force is entering a robotic future - "Tyndall AFB in Florida is the first base to deploy robotic patrol dogs."
New White House guidance downplays important AI harms - "Following a February 2019 executive order, the U.S. Office of Management and Budget (OMB) issued its final guidance on the regulation of artificial intelligence (AI) on November 17, 2020. This document presents the U.S."
Uni revealed it killed off its PhD-applicant screening AI - just as its inventors gave a lecture about the tech - "Fears of bias put compsci dept into damage-limitation mode after years of using it to analyze applications"
Google Chief Apologizes for A.I. Researcher's Dismissal - "The researcher, one of the company's best-known Black female employees, said she was fired last week."
Timnit Gebru: Google's "dehumanizing" memo paints me as an angry Black woman - "In an interview with VentureBeat on Wednesday, Gebru talked about toxic workplace culture in Big Tech, ethics washing, keeping AI research free from corporate influence, and the two things she wants young women and people of color interested in entering the field of machine learning to know."
The Dark Side of Big Tech's Funding for AI Research - "Timnit Gebru’s exit from Google is a powerful reminder of how thoroughly companies dominate the field, with the biggest computers and the most resources"
Expert Opinions & Discussion within the field
On the Moral Collapse of AI Ethics - Innovation and technological progress, in the context of racial capitalism, are predicated on profound extraction so the next disaster is inevitable, our failure to respond and offer an alternative worldview is not. We can do better."
Give the A.I. Economy a Human Touch - "Embracing artificial intelligence can help us create a new, equitable social contract - but only if we remember what makes us human. "
That's all for this week! If you are not subscribed and liked this, feel free to subscribe below!
Copyright © 2020 Skynet Today, All rights reserved.