The brain’s wiring is constantly evolving. Neural pathways are reshaped as we engage with the world and acquire new knowledge. At MIT\u2019s McGovern Institute for Brain Research and York University, scientists are using brain activity analysis and computational modeling to understand these changes better.<\/p>\n
McGovern Institute postdoc Lynn S\u00f6rensen, MIT Professor James DiCarlo, and York University Assistant Professor Kohitij Kar collaborated to study how animals and an artificial neural network, designed to mimic brain architecture, learn to visually identify the same objects. As the model improved, it reorganized itself in a manner similar to changes observed in the animal brains.<\/p>\n
Their research, published on July 8 in Nature Communications, demonstrates how visual processing changes enable animals to learn to distinguish new objects. By modeling these changes, the team aims to predict how training alters perception, potentially influencing educational strategies for diverse learners.<\/p>\n
When learning about a new object, multiple brain areas are involved. Visual-processing regions interpret information from the eyes and communicate with other brain parts to provide meaning and guide actions. The researchers sought to understand how these changes are distributed during learning.<\/p>\n
There has been debate over how much the brain’s visual-processing areas change when new objects are recognized. Some believe these pathways remain stable to maintain visual perception, while others report activity changes in visual-processing areas with such learning.<\/p>\n
The team focused on the inferior temporal (IT) cortex, a crucial component of the visual object-processing network. By the time information reaches the IT cortex, object features are so well-represented that it\u2019s possible to decode what object the subject is seeing.<\/p>\n
They recorded neural activity in the IT cortex of animals viewing and recognizing object images. Some animals were untrained, while others had learned to identify similar objects, allowing them to distinguish items like elephants and chairs, even in varying conditions.<\/p>\n
The broad IT cortex activity pattern was similar in both trained and untrained animals, indicating learning did not drastically alter this high-level visual representation. However, subtle differences were noted in how neurons responded to images in trained animals compared to untrained ones.<\/p>\n
The researchers used computational models to explore how these modest changes might impact learning. S\u00f6rensen trained artificial neural networks, mapped to the IT cortex, to identify the same object categories the animals saw, using gradient descent to improve accuracy.<\/p>\n
Only some models exhibited learning behavior akin to the subjects. In these models, changes in the IT-like stage mirrored learning-related changes in the trained animals\u2019 IT cortex.<\/p>\n
Though gradient descent is a common AI training method, it’s not seen as a direct model of brain learning. The strong learning effect match between the animals and the model suggests these neural networks can provide insights into biological learning.<\/p>\n
\u201cThis shows that you can actually build in silico versions of future experiments,\u201d S\u00f6rensen says. \u201cI think that gives us this playground of asking \u2018what if\u2019 questions \u2014 and potentially predicting new things that go beyond the experimenter\u2019s intuition.\u201d<\/p>\n
Most changes facilitating learning in the model occurred outside the IT cortex. \u201cThis tells us that a lot needs to change between the area we recorded from and the final behavioral readout during this process,\u201d Kar notes, highlighting the model\u2019s utility for exploring how downstream brain areas contribute to learning.<\/p>\n
The researchers emphasize their study’s granularity in measuring brain activity, which may not be possible in human studies, and note that animal brain organization has direct relevance to human learning. Understanding IT cortex plasticity could help design human learning strategies.<\/p>\n
\u201cOur prior conceptual working model of you learning new objects was that your brain makes changes to synaptic connections that are largely downstream of your visual system, so you don\u2019t destroy your visual system,\u201d DiCarlo explains. \u201cBut this study showed that your IT does change a bit to make it more relevant to new objects.\u201d<\/p>\n
These changes may impact recognizing other visual features, making it easier to identify objects beyond the learned ones, DiCarlo adds. Computational modeling can reveal these consequences, as demonstrated by the team\u2019s models showing increased information about object locations in the IT cortex after learning.<\/p>\n