Find your next reward
Many of our actions are determined by rewards, from opening a cold drink at the end of a long day to receiving a compliment from a colleague. However, the same systems that motivate behavior can be overdriven by addictive drugs or weakened by disorders such as depression. These characteristics of reward systems are what Assistant Professor Emily Silvestruck is studying.
“In my lab, we think a lot about how to set expectations and evaluate results: ‘Will this cake live up to the waiter’s hype? Will the movie live up to the trailer?'” Silvestruck said. “After learning how strongly the brain responds to disappointment, setting expectations properly has become something of an occupational hazard.”
In her research, Silvestruck investigates which neurons are activated during reward-seeking behaviors such as eating, drinking, and socializing, and how they work together to evaluate outcomes. Similar to Neal’s process, she uses AI to automatically track and label the mouse’s actions and facial expressions, synchronizing them with recorded brain activity.
“Knowing which brain cell types to target is important, because if you’re trying to develop drugs to treat neuropsychiatric diseases, you need to know which knobs to turn,” Silvestruck said.
For many years, neuroscience experiments in this field were limited to tightly controlled conditions, such as pressing a lever to obtain a reward. That’s because analyzing all unconstrained behavior requires datasets that are too large to analyze manually.
AI allows researchers to measure that complexity rather than filter it.
“The great thing about these tools is that versatility is now a feature, not a bug or a limitation,” Sylwestrak says.
Although AI can identify interesting behavioral motifs, Silvestruck emphasized that scientific intuition remains important.
“AI cannot completely replace curious and excited researchers,” she said. “While the process is enhanced, machine learning-based output requires human interaction. I don’t think researchers’ own curiosity and observation skills are obsolete.”
And when it comes to AI apps like chatbots that generate content, Silvestruck reminds students and the next generation of scientists that those tools are designed to find patterns in existing text and predict the words that are most likely to follow.
“In science, we don’t want to do the next most likely experiment. We want to do the next most interesting, most fruitful, or most creative experiment,” she said. “If you let AI do everything, it becomes derivative and not transformative.”
