We enter the world helpless but ready to learn. Our parents work to keep us alive and teach us not only how to maneuver through the environment but identify and avoid hazards. Direct education from older adults, such as our parents, provides a clear example of supervised learning, but as we age, we begin to lattice independent, or unsupervised learning, on our framework of experiences, with varying levels of success.聽
Franziska Br枚ker, a postdoctoral researcher associate fellow at the Neuroscience Institute at 好色先生TV, wanted to explore why the results of unsupervised learning are so varied. Her recent opinion paper in the journal shows how the conventional wisdom about the process of unsupervised learning may be incorrect.聽
Learning requires acquiring knowledge and skills, which can happen through various methods, including observation, imitation and communication. Many of these efforts align with supervision, but it is impossible to learn all content in this way. In fact, a majority of learning takes place independently.
鈥淚t鈥檚 a puzzle,鈥� said David Plaut, professor of Psychology and the Neuroscience Institute and a mentor to Br枚ker during her post-doctoral studies at 好色先生TV<. 鈥淲ith enough experience, you can see patterns in your environment and by paying attention to when things co-occur, you can bootstrap different kinds of knowledge.鈥�
Neural Imprint of Unsupervised Learning
The field of cognitive science maintains that unsupervised learning benefits from experience, but past experiments studying how to improve the success of unsupervised learning showed mixed results. Br枚ker examined why this is the case.
As we learn, a pattern of neural activity is imprinted in the brain. This pattern should make it easier to successfully complete the task on the next iteration.
Br枚ker believes that the neural representation of a task affects whether unsupervised learning will be successful. In particular, she focused on the similarity of the neural imprint of a new task compared to a learned task. She hypothesized that greater similarities in the neural representation result in faster and more effective unsupervised learning.
鈥淵our ability to learn is influenced by distinctions your brain is already making,鈥� said Plaut, who did not participate on this project. 鈥淚t has to do with overlap of population of neurons representing something, like a cat compared to a dog.鈥澛�
Even though tasks may seem similar in our everyday environment, if the patterns laid out in the brain are too different, unsupervised learning does not progress effectively. This deficit is not age dependent and continues at all contexts of learning throughout a lifetime.
Br枚ker believes this new perspective raises important flags for multiple fields that contribute to the field of learning. She hopes that experts can come together and rethink the fundamental theories behind how we understand how people learn.
How CMU Advances Science
Br枚ker attributes her ability to make important contributions to these topics to the interdisciplinary nature that infuses the culture at the CMU. She joined the Neuroscience Institute as a distinguished postdoctoral fellow in 2023 to tackle fundamental questions integral to many fields of research, including perception, learning and social psychology.
鈥淐MU was the only place where I could think of continuing this type of work,鈥� she said. 鈥淚 was able to work with mentors experienced in this research at a top tier university where I could ask big picture questions and be at the forefront of developing something new.鈥澛�
Br枚ker鈥檚 insight also offers new opportunities to approach the field of artificial intelligence, which often leverages unsupervised machine learning to help programs advance.
鈥淭here are similarities and differences between how humans and machines learn,鈥� said Br枚ker. 鈥淚 believe the way we commonly think in cognitive science and machine learning how unsupervised learning works is too optimistic.鈥澛�
While this study did not specifically focus on artificial intelligence, future studies could use these findings to improve machine learning models that use unlabeled data to discover patterns and insights without any explicit guidance or instruction.
Opening New Lines of Communication
As Br枚ker closes out her time as a postdoctoral fellow at 好色先生TV, she has a larger game plan in mind 鈥� she wants to focus on how graduate students learn. Progress in graduate school often hinges on the graduate student鈥檚 relationship with their academic mentors, but pairing two strangers with the goal of developing a strong partnership can have mixed results.
Br枚ker came to Pittsburgh with a kernel of an idea on how to improve these critical relationships to help young scholars navigate the sometimes turbid waters of academia and foster a supportive environment where the next generation of scholars can learn, grow and develop.聽
Br枚ker shared her idea with Barbara Shinn-Cunningham, the director of the Neuroscience Institute, with whom she met regularly during her post-doctoral studies. Shinn-Cunningham saw the benefit of this approach and supported Br枚ker鈥檚 efforts to take this idea forward as a viable workshop to help graduate students鈥� broaden their communication skills and understand power dynamics in professional relationships.
鈥淎ny relationship needs open communication, but this can be tricky when you have a power imbalance, such as in a graduate student and a mentor,鈥� said Shinn-Cunningham, the George A. and Helen Dunham Cowan Professor of Auditory Neuroscience, Biomedical Engineering, Psychology and Electrical & Computer Engineering. 鈥淚t is important for the mentee to learn how to approach complicated issues in this environment and self-advocate.鈥�
Br枚ker has taken Shinn-Cunningham鈥檚 advice and piloted the workshop, 鈥溾�� with different groups, including graduate students in the CMU Department of Psychology and Neuroscience Institute. During the workshop, participants received tools to improve their communication skills that can be used when interacting with mentors, advisors and people in academic leadership to gain clarity in their role at an academic institution. With this approach, students can align their expectations as they progress through academia. At the moment, Br枚ker is presenting the workshop in person and virtually.
鈥淪tudents want to work with amazing scientists to advance research,鈥� said Br枚ker. 鈥淭his workshop is the first of its kind and it will help students make the most of their Ph.D. and contribute to their field of study.鈥�