b'PARTNER CONTENTWhy Human Expertise Still Matters in the Age of AINew technology means seed testing is more precise than ever, but the expertise ofthe people in the lab makes all the difference.T heres a lot of talk these days about howof features, quality of data collected, algorithm com-technologynamely multispectral imaging,plexity and so on.artificial intelligence (AI) and associatedThat is where human expertise comes into play. advancementsis changing the world of seedNo matter how well the model performs, it lacks one analysis.crucial thing: common sense. It is simply evaluating Our team is increasingly working with a tool thatpatterns and data against the dataset it has been is becoming more common in seed testing, and thatgiven as a training set. is multispectral imaging. It allows for a deeper, moreIf it predicts that a seed is of a particular species, comprehensive examination of seeds, far beyondthe prediction is just a numberit does not under-what the human eye can perceive.stand environmental factors that need to be taken By capturing images across multiple wavelengthsinto account nor the real-world implications of its of light, from visible to near-infrared, multispectralguess, unless specifically trained as such. It cannot By Janek Bartel, imaging can reveal critical information about seedidentify something correctly if it has never seen it Lab Manager, SGShealth, quality, and composition that traditionalbefore. It lacks context, something only a human can Canada Crop Science methods might miss. It has completely changed whatprovide.we are able to do for our customers. Yes, the technology is advancing rapidly, and yes, That said, when people talk about these newit is incredibly powerful. But without a human at the technologies in seed analysis, I notice how quicklycontrolswithout the ability to see beyond the basic the conversation can become convoluted. Despitedata and what it really signifiesAI used in seed all this advanced technology, we still need talented,analysis as we know it today will always need human knowledgeable human experts in the loopandoversight. that will not change anytime soonif ever. You still need an experienced analyst to look at Imaging tools of all kinds are often found inthe seeds, verify predictions, and provide the context seed labs these days, and the data they produce isthat machines simply cant. It is not replacing jobsincreasingly combined with some sort of algorithm.it is evolving the way we work.These models are limited by the data they are trainedWe are in a phase where AI and human analysts on. We are far from the point where they can figureneed to work hand-in-hand. The machine learning things out on their own, especially in niche fields likemodels are powerful, but they are not all-knowing. seed analysis.And while it might be tempting to think we can auto-Right now, we are in the Wild West of AI. The ideamate everything, human judgment is still irreplace-of AI spontaneously developing the intelligence toable. These models can only be as good as the data sort seeds autonomously is not here yet, becausethey are trained on, and for niche industries like seed we need high quality data and someone to train theanalysis, we just do not have the vast, good quality model. and easily accessible datasets needed for AI to learn I have worked directly on some models, trainingunsupervised.them to sort seeds and predict characteristics likeAs exciting as AIs potential is, and as important purity, seed size, optical weight and even health. Theas its becoming in everyday seed analysis, there is process involves feeding the model thousands ofa balance that needs to be struck. Yes, we should images, or blobs, to train it.embrace these technologies and push them to their I recently spent 30 hours identifying featureslimits, which our team as well as other international and classifying just a few hundred blobs to train alabs are doing. But we should also remember that demo model. Now imagine needing to classify tensthey are tools, and all tools require a human who of thousands of blobs for a more accurate prediction,knows how to use them properly and when to use which often is the case. The time and effort requiredthem appropriately.are staggering. And even after all that, the modelThe analyststhe people with the experience might still struggle to identify simple differencesand intuition to interpret the dataare essential. between seeds, which then calls for a re-evaluationAnd that is not going to change.NOVEMBER 2024 SEEDWORLD.COM/CANADA 11'