In the second week of class, Yuyan Wang had about 35 MBA students experiment with the finer points of tuning neural networks, looking inside a model similar to those that make recommendations on platforms like YouTube and Amazon.
“This is art, not science,” she said as students tinkered with the algorithmic composition of simulated classification problems (distinguishing between cats and dogs, for example). For most people, it was probably their first experience adjusting the settings that control how a machine learning model “learns” – how it optimizes itself based on data to improve predictions.
“We don’t know what neural network configuration is best for our data set,” Wang explains. “This is why it’s called hyperparameters tuning. It’s basically a bit of trial and error, where you have to try out a few values and then look at how your model performs in your predictive scenario to see which type of parameter configuration gives you the best performance. ”
Wang, an assistant professor of marketing, began teaching. Understanding AI technology for business problems — One of the school’s first AI technology courses — January 2025. Before joining the Stanford Graduate School of Business in 2023, she spent nearly seven years in industry, working as a machine learning scientist and engineer at Uber and Google’s DeepMind.
quotation
The biggest challenge when launching a project or influencing a real-world product is actually not technical, but rather communication.
Author name
— Wang Yuyan
That experience has influenced her approach to materials. “We’ve found that the biggest challenge in launching a project or having any impact on a real product is actually not technical, but communication: convincing leadership, operations teams, or product teams that this is the right thing to do,” she says. “We speak different languages when it comes to AI.”
Wang’s goal is not to immerse students in the intricacies of algorithm design, but to develop future leaders who are well-versed in the vocabulary of AI and able to understand the potential of agentic AI. She utilizes a variety of formats, from lectures and mini-case studies to visits to practitioners. Jascha Sohl-Dickstein of Anthropic, the inventor of a model that powers image generation, visited her classroom during the winter semester, as did experts from OpenAI and startups like Neo4j and LangChain. Practical lessons help MBAs better understand the basics.
“I had a question about hyperparameters for model building,” said one student, asking about the number of layers of artificial neurons that make up the model. “For example, how do you decide whether to have just one neuron-limited layer or about 100?”
“That’s a great question,” Wang replied. “As a general rule of thumb, the larger the dataset and the more dimensionality of the input, the larger the neural network should be.” The ideal size of a neural network depends on how it will be used, she stressed. “You can’t just throw data at it, calculate a problem, and expect it to magically spit out a model that maximizes your goal.”
That day’s homework will include additional algorithm training using a larger dataset.
“The hardest part of applying AI to solve business problems is not building fancy models, but transforming business goals,” Wang says. “The translation part of that is not something that engineers can do. But if you only have business insight, it’s often not very useful. I want my students to have the ability to translate business problems into something that AI can solve.”
