Hello!

April 22nd was Earth Day, in case your Instagram feed being filled with nature pictures wasn’t enough to catch on. In honor of this, we've focused HTC12 on AI environmental-ethics-related stories — covering various ways that AI can both help, and harm, our planet.

(We considered recognizing a different kind of "green" day that was celebrated last week, but there was a disappointing lack of coverage).

This one goes out to you, mother Earth.

Reduce, Reuse, Recycle (with AI)

Industrial waste makes up at least 50% of all global waste. This amount is roughly 18 times greater than the amount generated by municipal solid waste (aka “trash”, if you don’t have much time). The complexity of the manufacturing process as well as the global supply chain leads to a lot of waste produced by inefficiencies, mistakes, or a lack of responsiveness.

How AI can help

AI algorithms can be applied to making the manufacturing and distribution process more responsive, adaptive, connected, and efficient, allowing businesses to save money and the environment. Algorithms can be deployed in factories to avoid and automatically respond to costly, waste-generating mistakes, in business to predict supply and demand for various products, and in complex decision-making to help companies find and respond to patterns.

Companies that care

Many big-name companies are working to reduce their environmental impact, including Subaru, Nike, Ford, and Coca-Cola, and AI is already playing a role. French food maker, Danone, was able to use AI to achieve a 20% reduction in forecast error and a 30% reduction in lost sales by using machine learning to predict demand for their products, according to Capgemini.

Computing Costs

AI algorithms require a vast amount of computing power to train. This power translates to massive amounts of electricity that go into the development of these algorithms. A paper out of UC Berkeley and Google estimated the environmental impact of the following algorithms:

• GPT-3, from OpenAI, produced 552 metric tons of carbon dioxide during its training (the same amount produced by roughly 120 passenger cars over a year).

• Meena, from Google, consumed 96 metrics tonnes of carbon dioxide (the same as power 17 homes for a year).

Minimizing impact

Google found that environmental impacts can be minimized when the training of these algorithms is optimized accordingly.

• Many of these high-powered AI models use neural networks (computing technology inspired by the human brain where artificial “neurons” form connections with each other to perform computations and recognize patterns). If these algorithms are designed to be sparse (meaning that each “neuron” only connects to a relatively small number of other “neurons”), this can increase the efficiency of an algorithm and reduce energy consumption by a factor of 10.

• Using computer chips specifically designed for neural networks can cut energy consumption by up to 5 times.

• Moving the training from a data center in a location where the primary source of energy is coal, like India, to a location that relies more on renewable sources, like Finland, can reduce consumption by a factor of 10 to 100.

Office Politics

Environmental impact was one of the concerns brought forth by Google’s AI ethics research team prior to the departure of Timnit Gebru and Margaret Mitchell. Jeff Dean, a senior executive vice president at Google, is a co-author of this paper and one of the people criticized for their role in forcing Gebru from the company. This paper also underestimates the impact of AI development on the environment when compared to other recent studies.

How Can AI Address Climate Change?

A recently published Forbes article outlines several use-cases for how artificial intelligence can be used to advance climate solutions. The authors even claim that "AI will be a major enabler at the core of climate change technologies."

Here are a few ways AI can help.

1) AI for net-zero waste

Predictive modeling algorithms can be used to better optimize resource allocation.

• Take, for instance, Wasteless — a startup that sells AI-powered computer vision software to supermarkets and grocery stores to help them recapture the full value of their perishable products and reduce food waste through AI-powered dynamic pricing.

2) AI for environmental intelligence: better climate predictions

Environmental Intelligence has become especially significant for predicting climate trends and patterns that are extreme in nature. Artificial intelligence and deep learning have the ability to quickly analyze dynamic systems and simulate them, producing accurate models that can then be applied by scientists and researchers for more robust decision-making.

• Initiatives like the Ocean Data Alliance have implemented AI to develop comparative Smart Ocean City Action Plans, leveraging data to guide policy recommendations.

3) AI for environmental monitoring (drones!)

Another area where AI is making an impact on climate change is with drone surveillance of forests. Technology entrepreneur Ewan Kirk says that "drones are one of the easiest ways to collect important data from remote regions, and gather intelligence on ecological health."

• For example, a Finnish startup, Aeromon, leverages drones to track industrial emissions in real-time “Aeromon reveals the true extent of airborne emissions, enabling automatic reporting of emissions with a 360-degree view giving our customers real-time process insights.”

4) AI for food supply chain optimization

AI is becoming prevalent in the food supply chain industry from better predicting demand in restaurants, reducing food waste, and even helping developing world farmers diagnose and treat agricultural crops.

• Indian farmers have been able to achieve up to 30% higher yields with machine learning advice on when is the best time to sow crops.

TL;DR: AI can have really awesome impacts on the ~greatest existential threat~ facing our generation. Here at RAISO, we're eager to continue learning about different use-cases, seeing first-hand how AI can solve some of the world's most complicated challenges.

Weekly Feature: The Future of Reinforcement Learning aka the Self-Taught AI

If you’ve ever tried to train a puppy or taken an Intro to Psych class, you already know the basics of reinforcement learning. It’s the process of learning in an environment that gives feedback or a reward based on behavior, like giving a dog treats bit-by-bit until they learn to roll over. Now imagine AI using this model to optimize businesses.

How is this different from other machine learning techniques?

When we think of machine learning, we often think about supervised learning, the most common way that AI is taught. In supervised learning, AI relies on a supervisor to give them guidelines and verify their algorithmic predictions.

• For example: an AI might predict that the weather in Evanston will be 45℉ today based on the data from the past month, but if we had a crazy April heatwave that bumped the temperature to 65℉, the AI would learn to realign its next prediction to be more accurate and grounded in truth.

• This is the type of AI that many businesses use today.

With reinforcement learning, there is no truth or guideline to follow. In this case, the supervisor doesn’t give them the “right answer” because there isn’t one – the task is too complex. AIs learn by trial and error, doing rather than predicting. So supervisors give the AI a reward if they complete the task, then ask the AI how they got to that goal. From there, they assign value to the right actions.

• For example: if the AI was learning how to walk, the supervisor would give them a reward when they stood up. The AI would then continue interacting with their environment and the supervisor would keep giving them rewards that would get the AI closest to walking, or the most efficient actions.

• This is how gaming AIs like Deepmind’s AlphaGo can become better at games than humans – they learn on their own based on learned strategies rather than regurgitating formulas.

How is this applied in the real world?

With reinforcement learning, AIs don’t need heaps of historical data to rely on, they can learn while they go. This helps automate and optimize processes much quicker than ever before. But the biggest value of reinforcement learning is that they can learn complex tasks in creative ways.

Highlighted in the Harvard Business Review, the technology is now being implemented into business settings. Google uses reinforcement-learned AI to control cooling in their data centers. Spotify uses it to make suggestions for your playlists.

Kathryn Hume and Matthew E. Taylor, head of Borealis AI and associate professor of computing science at the University of Alberta, respectfully, say that businesses should be utilizing this technology in situations where rules and formulas just won’t cut it. If there are tasks that are too dynamic or have too many exceptions, like how to sell stock in small enough orders that minimize its price drop, reinforcement-learned AI can be the solution. Though it may be a large investment, it may be worth it in the long run to outshine competitors.

Read the full story here and watch a video about reinforcement learning here.

Written by Sophie Lamb, Molly Pribble, and Lex Verb