It’s seems like artificial intelligence (AI) exploded onto the marketplace overnight. And then began hitting at agtech pretty aggressively.
But, as presenter Ron Green, co-founder and chief technology officer of KUNGFU.AI presented at Potato Expo 2024, it has taken humanity about 75 years to get to this point in AI technology. And even until now, there have been multiple hype cycles.
“Each time was kind of over promised and underdelivered,” said Green during the presentation, which makes now different and a very exciting time to be involved in this technology.
There are two approaches to AI that have dominated the tech: symbolic AI and connectionist AI.
Symbolic AI is the traditional approach, and is based on the concept that many aspects of intelligence can be achieved by manipulating symbols to form logical statements or rules.
Connectionist AI is what the general public tends to imagine when it hears the term “artificial intelligence”. It parallels the way people understand human brains to work. Like connections made among neurons in the brain when humans learn, similar connections are made via artificial intelligence when the machine learns from data. This allows it to make decisions or predictions. This is the dominate approach used today.
Artificial intelligence is a blanket term for the tech behind building a program that can learn, proceed, and overall has capabilities similar to humans. Beneath that, as a subset of AI, is machine learning, which is a way to achieve artificial intelligence.
Farmers are often familiar with this in contexts where machines are fed tens of thousands of images (or other data) — say, of potato pests or leaf diseases — from which a system can learn and make decisions. Similar to this is deep learning (which dominates machine learning), which uses a human-like “brain” approach by learning from mass amounts of data, often numbering in millions.
Green mentioned six key developments that have taken shape over the last decade to make possible the AI that everyone is familiar with today.
- An explosion in available data. “Deep learning systems require an enormous amount of data to work, and right when the system sort of needed the data, the data became available,” said Green.
- Hardware innovations. With this amount of data, a massive amount of computing power is required. As luck would have it, the computer processing units (CPUs) used for computer gaming are perfect for the type of math needed for deep learning systems. “We would not be where we are today had all the advancements in computer gaming not happened in early 2000s,” Green said.
- Algorithmic advances. While the theory of AI goes back many decades, most of what happens today is built on what was discovered under 40 years ago. But those systems also couldn’t handle the large amounts of data.
- Open source. Communities that allowed for open source software have been fundamental to the development of AI. Open source means the software can be freely used, distributed, edited, and shared. “All AI today is built on pretty much, without exception, one or more … technologies that are…completely free,” said Green. It has allowed the democratization of AI, allowing anyone to participate.
- Transfer learning. These systems can be taught to do one task, and then use that learned task and apply it to a different and related task. This means that no one is truly starting from scratch, and everyone who pushes AI forward will help their successors continue to push AI further.
- Supervised learning. Systems can pre-train on a task that has nothing to do with its future use, but it builds a foundation.
Green warned against treating AI as traditional, conventional software. It isn’t, and treating it as such often leads artificial intelligence initiatives to early failures. There’s also often a misalignment in terms of ROI. AI can be costly, so it’s important to understand the ROI on any projects built on using AI. Data needs are often underestimated. All modern systems require training on mass amounts of good quality data, and the absence of that could also result in a project’s failure.
“I would say be careful about over hyping expectations. Modern AI systems are incredibly powerful, but they’re not omnipotent. They’re not omniscient,” said Green. “So go in eyes wide open and understanding what they’re capable of with the right expectations.”
Current use cases in potatoes include identifying and addressing potato-related misinformation posted online by posting factual corrections, use in drones for land classification, pest and disease infestation monitoring and management, supply chain optimization, automation of machinery, and much more.
Other Articles
Pecan scab resistance: Three varieties showing promise
The Element – Aigen’s autonomous robot powered by renewable energy
