Dr Martin Haesemeyer | A Deep-learning Framework that Links Brain Activity with Behavior

Deep-learning algorithms are computational tools that can rapidly analyze large amounts of data to detect patterns and make predictions based on these patterns. Convolutional neural networks – or CNNs – are a type of deep-learning algorithms used to analyze images.

Dr. Martin Haesemeyer at the Ohio State University and collaborators recently developed a new CNN-based framework that can unveil hidden links between neural activity and animal or human behavior, by analyzing neuroscience datasets. Their framework is called MINE: Model Identification of Neural Encoding.

Most conventional approaches first decide which relationship to look for in the data and then identify patterns that fit this relationship. MINE takes an opposite approach, as it first trains a flexible CNN and then analyzes the learned model to unveil relationships between brain activity and predictors, such as external stimuli and a person’s internal states and behaviors. MINE then characterizes these observed relationships. This approach gives MINE the flexibility to process different types of imaging and experimental data, without the need to transform this data beforehand.

To test their framework’s performance, Haesemeyer and his collaborators used it to analyze an experimental dataset containing neural activity across the mouse cortex and corresponding behavioral responses.

They also used it to interpret data they collected during experiments on zebrafish larvae. These experiments aimed to identify neural circuits that allow zebrafish to regulate their body temperature. The researchers changed the temperature of larval zebrafish under a microscope while recording their tail movements and neural activity.

The MINE framework allowed the team to characterize neurons based on their receptive field and how complex their computations are. In addition, the researchers identified neurons that link the temperature that zebrafish sense with their own ongoing behavior. Notably, conventional statistical methods such as clustering and regression had never uncovered this type of neuron.

The team’s results highlight the value of the MINE framework for exploring the complex relationships between neural activity and behavior. When used to analyze human brain activity, MINE could reveal new insights into the functions of different brain regions, leading to a deeper understanding of various psychological conditions.