Hold the Code #19
From point A to point B
Hello!
Welcome to Hold The Code #19 and our first summer edition! 🏖
We're very excited to continue sharing news about AI this summer with our readers from NU and beyond.
In this edition, we cover a recent tech bill passed by the Senate, how AI is changing the postal service, ways that AI can alter the product design process, and a new essay that argues reinforcement learning can help AI achieve general intelligence.
See you next week!
Senate Passes $250 Billion Bill to Boost Tech Research
Last Tuesday, the Senate approved a bipartisan, $250 billion bill boosting government spending on technology research in light of rising competition from China and other countries.
According to the Wall Street Journal, "The legislation represents a potential landmark effort to turn the tide on several long-term trends in U.S. competitiveness. Those include eroding federal investments in research overall and a shrinking share of the world’s semiconductor manufacturing."
Although the bill still needs approval in the House before it can end up on President Biden's desk, the bill's wide support in the Senate indicates that passage is likely.
What does the bill do?
The bill aims to overhaul U.S. government support for science by expanding the government’s role in technological research, including through the National Science Foundation.
It would authorize about $190 billion in spending to strengthen U.S. advanced technologies to better compete globally.
The bill aims to become a vehicle for a number of other tech-related initiatives. One would authorize about $52 billion for encouraging more semiconductor production in the U.S.
It provides $100 billion over five years for a new technology directorate at the NSF to fund research into artificial intelligence and machine learning, robotics, high-performance computing, and other advanced technologies.
An additional $10 billion would be authorized for the Commerce Department to designate regional technology hubs for research, development, and manufacturing of key technologies.
The House committee plans to vote on its research package next Tuesday, and the Biden administration said recently that it supported the Senate bill and would continue to work with Congress to improve it. Overall, many are optimistic that “as our nation works to recover from the worst economic and public health crises of our lifetimes, now is the time to make these major investments in our long-term economic resilience and competitiveness.”
You've Got M-AI-L
To help process 7.3 billion packages a year – 231 per second – the U.S. Postal Service has begun using artificial intelligence.
Earlier this month, the USPS implemented a new strategy that involved using advanced computer systems at 195 mail processing centers nationwide. These centers are able to apply different computer vision models that function to reduce the time it takes to track lost packages, among other tasks. Some estimate that certain tasks that otherwise would've taken several days can now be accomplished in less than 2 hours using these programs.
How does it work?
The agency has begun using Edge Computing Infrastructure Program (ECIP), a distributed AI system running on the NVIDIA EGX platform. Each edge server processes 20 terabytes of images a day from more than 1,000 mail processing machines. The Triton Inference Server, the open-source software from NVIDIA, provides the AI models each of the mail processing centers needs.
“The overall design here is to continue to enhance and build a database for packages so that they can over time improve package processing and efficiency and build from this model,” Anthony Robbins, the head of public sector relations at NVIDIA, said.
Future ambitions
Though long inhabited by funding and operations, the USPS has bold goals for future AI applications. For instance, it’s looking into using autonomous vehicles for mail delivery and monetizing its geolocation data. A February Information Technology and Innovation Foundation report touted the use of robotics for last-mile postal delivery.
Designed by AI
What do cars, computer chips, and Nutella jars have in common? They are all being designed by AI. A working paper out of Harvard Business School argues that AI can “profoundly change the practice of design” by automating the learning process to produce designs for new products, changing the way the design process works, and changing the roles of designers themselves.
Who's using it
Renault is using AI to simulate the predicted behavior of automated manual transmission systems in their cars.
Google is using AI to develop their next generation of tensor processing units (TPUs - computer chips that have been optimized for use in neural networks, a machine learning technique).
Even Nutella is using AI to design new packaging on their “Nutella Unica” jars.
Ethical AI for ethical design?
The Harvard paper raises interesting points regarding the positive benefits of AI involvement in the design process, including reduced cost, increased design iterations, and (arguably) a more user-centered design approach. However, one of the main case studies they examine is Netflix’s algorithm for designing movie posters based on user data, an algorithm that was criticized for targeting users based on race and misrepresenting the roles of Black actors in certain TV shows.
AI seems to be a fascinating new development in the design space - especially in terms of optimizing business cost and investment, but this makes the design of these algorithms all the more important as their impact in the field of design grows. By giving the power of design to computers, we need to place more of an emphasis on how these algorithms are designed, for whom, by whom, and for what purpose.
Weekly Feature: Recreating Intelligence with Reinforcement Learning
DeepMind's researchers believe reinforcement learning is the key to recreating general intelligence in artificial systems
Winter is coming. A squirrel leaps from branch to branch in search of nuts. Revving up her motor and sensory engines, she looks for tree nuts while keeping an eye out for potential dangers. Strong skills will reward her with delicious nuts and survival, but if her skills fail her, a mundane task can turn into a fatal leap. Depending on her actions, the squirrel is either rewarded or penalized in terms of her chances for survival. With each success and misstep, the squirrel becomes better at maximizing rewards while avoiding penalties, adopting the best practices to achieve her goal - collecting delicious nuts to last through the winter.
This process - learning how to maximize reward - helps organisms evolve general intelligence, which is essentially general problem-solving abilities through coordinating multiple skills (this is contrary to narrow intelligence, which describes an ability to perform one specific task, such as speaking). Currently, AI systems that target artificial narrow intelligence (such as computer vision, NLP, etc.) are blooming, but some researchers at DeepMind are captivated by a different approach to cracking artificial intelligence. They are interested in accelerating artificial general intelligence through recreating the evolution of natural intelligence - particularly, through applying the simple yet effective mechanism of reward maximization that gave rise to natural intelligence. Reinforcement learning, they find, can effectively replicate this process of developing intelligence in artificial software systems.
Reinforcement learning is a type of AI algorithm that consists of 4 key elements:
an agent
an environment for those agents to act in
a set of actions that the agent can take
rewards/punishments the agent receives for taking those actions
Let's break down how these elements work together using the squirrel example. The agent (squirrel) performs actions in the environment (tree), and based on how effectively the action moves the agent towards its goal (collecting nuts), the agent is either rewarded or penalized (survival). The agent often knows nothing about how to navigate the environment initially, but through the feedback from hundreds of rewards and penalties, the agent learns how to maximize rewards in this environment and is now "intelligent".
The goal is to use reinforcement learning to create intelligent systems that can piece together multiple intelligence factors such as perception, communication, social intelligence, and so on to solve complex problems. There is still a long way to go for reinforcement algorithms and general intelligence. While researchers continue to combat the various challenges to achieve the goal, it's something we can look forward to.
Read the full article here.
Written by Larina Chen, Molly Pribble, and Lex Verb