Smart Sorters Part 2:
How Does a Sorter Learn?

In order to create a smarter sorter, we studied in depth how a sorter learns. First of all, let’s address the elephant in the room: What difference does it make how a sorter learns? After all, if a sorter is doing its job, does it matter how it accomplishes it? Yes, results matter but so does efficiency; some learning algorithms are more optimized for sorting than others. This can mean the difference between using a computer to control a single lane of a sorter vs using the same computer to control multiple lanes.

"The sorters must be able to generalize the defects’ appearance and distinguish them from normal features like the blossoms and stems."

So how does a sorter learn? Like human sorters, vision sorters learn how to measure size and color and identify defects through examples. These examples are typically photographic images of defects that must be sorted out. The sorters must be able to generalize the defects’ appearance and distinguish them from normal features like the blossoms and stems.

However, generalizing the defects doesn’t mean simplifying them. We wouldn't put deep neural networks in Sunsort^ai if we were just looking for brown spots on fruits. We've already written vision algorithms to do that.

In fact, we tried to build AI by integrating multiple vision algorithms. It is possible to have specialized algorithms: one detecting canker, another one detecting wind scar, and so on, then run these algorithms in succession, making several passes per image. It’s as if there are multiple specialists, each taking a turn looking at the same image. It worked, but it didn’t get us closer to a smarter sorter.

Having a multipass system is effectively having many human sorters, each specialized in recognizing one defect. However, there are many different defects and it’s not uncommon for the same defect to manifest itself differently from one region to another or from one season to the next. We knew that having hundreds of specialists inspecting one image of fruit in fractions of a second would require the computing power of leading edge data room servers which wasn’t practical for packhouses.

"With Google AI’s toolkits, we were able to build the next generation of Sunsort..."

We believe that simply increasing computing power is not the best solution. We need a sorter that can identify multiple defects and grade the fruit in a single pass, regardless of whether the fruit is flawless or has several defects. These are not trivial qualities that we are aiming to solve. Making a smarter sorter is not as simple as copying and pasting algorithms. So we had to shift our entire design paradigm. Soon after we started to build Sunsort^ai, we knew that we needed help. We researched and studied all the AI resources available and came across Google AI. With Google AI’s toolkits, we were able to build the next generation of Sunsort that learns and identifies multiple defects in one pass in a fraction of a second while running on our industrial computer.

When we reexamined how our machine learns, we saw that intelligence isn’t only defined by computing power, but how that power is used. We learned this lesson with the support of Google AI and the brilliant minds that created it.

How does your sorter learn? Does it feel more like a machine or human?