Machine learning has monopolized the dairy tech-talk spotlight for years — in fact, a 2020 article in the Journal of Dairy Science on precision dairy monitoring technologies (PDMT) noted that such data-analysis tools were being used on farms going back to at least 2017. But with the recent leaps-and-bounds growth in artificial intelligence (AI) models’ capacities to digest data and create predictive statistical algorithms, the attention has reached fever pitch. A recent Cornell University “Cow Convo” podcast episode explored the potentials, the constraints, and the current realities behind the hype with Miel Hostens, associate professor of digital dairy management and data analytics. The discussion on AI’s uses ranged from labor efficiency to body conditioning scores to privacy. Hostens said that despite the breakthroughs in machine learning techniques for tasks such as lame cow detection, body conditioning scoring, and other monitoring processes, there are some roadblocks.

“We need to see proof that these algorithms are really able to detect earlier than a human, and we haven’t seen that yet,” he said. Another potential drawback is overreliance on the monitoring software systems. “Those systems have to be available and running all the time. What if the hardware breaks down and there’s no one on farm who can detect labor anymore?” he asked. “And cameras get dirty; they need maintenance.” He said the technology needs to be consistent, reliable, and generalizable across farms.

During the podcast, Hostens also addressed an ancillary concern: the ethics of data collection and ownership. While “the one who can verify the truth of the data is the owner,” he said, it’s also true that two people can own it, which makes data different from other assets. Hostens pointed out that the farmer can delegate ownership to the AI company, which could then use it to train its algorithms. The ethics consideration has been a particularly thorny issue for AI tools which monitor human, rather than bovine, behavior.

“Many commercial farms are trying to find ways to do this. Monitors in milking parlors have been around for years,” he noted. However, in the past, this data was not typically stored, and certainly was not used to create algorithms.

Other AI tools in the works or already on the market monitor the herd, not the helpers, by looking for early signs of disease or estrus as well as taking measurements for classification. Holstein Canada recently launched a classification research program and has put out a call for participants. They aim to use 3D cameras during traditional classification visits to compare the system’s measurements to those collected by the humans. If the computer generates accurate information, then an algorithm could be built around collected data and the resulting system could augment, streamline, or even replace human-generated measurements.

Holstein Canada is not the first to explore classification software; in “Cow Convos,” Hostens noted that such systems exist already. “We are seeing some of the first signals that these techniques can work,” he said. But he warned that the old adage “garbage in, garbage out” applies, and farmers can help prepare for the AI transition by standardizing their past and current records. “The algorithm will only be as smart as the ground truth used to train it was valuable and correct,” he said.

To comment, email your remarks to intel@hoards.com.