Alex Senden, a second-year master’s student in the U of M’s department of computer science, is developing artificial intelligence (AI) tools that could help researchers distinguish crops from weeds in agricultural images.
“My research is in the domain of AI machine learning,” Senden said. “It’s basically computer vision, which is this idea that you’re trying to get computers to understand and make intelligent decisions about images or videos or visual data.”
Senden works on image generation, training models that learn to create realistic new images from random noise. “You basically start with a bunch of random noise, complete TV static, and the model […] learns to go from this noise to something that looks like a real image,” he explained. By feeding the model a bunch of real images during training, it learns to reconstruct order from randomness, turning static into a believable picture.
He is especially interested in models that not only create images but also know exactly what each pixel represents. “I want to be able to generate images where we know where things are in the images,” Senden said. “When I say that, I mean literally pixel by pixel, I can tell you what object this pixel is representing.”
This research approach is being applied in collaboration with the department of agriculture to develop crop field imagery. “They have this big problem right now where they want […] a machine learning model that can go through real images in real time and tell you, ‘Here’s where the weeds are, here’s where the crops are,’” Senden said. Beyond this, such an approach also holds broader potential for agricultural research as U of M collaborates closely with Agriculture and Agri-Food Canada.
The challenge, he added, is that such models need huge volumes of annotated data — images that researchers need to label manually, marking each pixel to show which areas represent crops and which represent weeds. This process is often slow and extremely expensive to carry out. “To do that for every single pixel in an image, it takes way too much time,” he said. “You can do it for small amounts of images, but to do it […] on the scale of a crop field [is] pretty infeasible.”
Senden’s long-term goal is to address this challenge by generating artificial images with built-in labels. With data-labeling automation, his system could provide the data needed to train models without relying on years of manual work. The same method could support other fields that depend on detailed imagery, such as medical diagnostics, where identifying tumors or lesions also requires large, labelled datasets.
“I would like […] to make this type of machine learning model or computer vision model […] easier to train so that it’s really easy to implement across industries or research,” he said. Beyond his own project, Senden follows emerging work in federated learning — a framework that lets multiple users train AI models collaboratively without sharing private data. “This has really big implications for data privacy,” he said. “Now we don’t actually have to share data in order to make this one model that has basically […] learned from all of our data.”
According to Senden, data privacy will become an increasingly important issue in the future, and federated learning might be able to offer a solution to one of AI’s biggest current challenges.
