Biophysics Seminar: Orbitofrontal ensembles integrate reward, movement, and taste predictions during learning, by Evan Hart
Monday, May 5, 4:00 pm
PSC 2136
Introduction
For my CPSG101 spring excursion, I chose to go to a biophysics seminar by Dr. Evan Hart. He is a psychologist working at the Brain and Behavior Institute. This week, he presented his research on brain maps using rats. Cognitive maps are very similar to regular maps. There are multiple ways to get from one place to another, and research has shown that rats have the ability to form these cognitive maps. His research was seeking to answer the questions "what info is in cognitive maps and why?" and "how does it change/evolve during learning?"
In order to answer these questions, he decided to look in the orbitofrontal cortex (OFC). He used a device surgically inserted onto the heads of rats to measure voltage spikes in the OFC. The characteristics of these voltage spikes - duration, frequency, amplitude - would tell him how many neurons were firing, and where they came from. Once he had that, he would measure the firing rate, in Hz, to see how activated these neurons were. The higher the firing rate, the most activated the neuron was. He designed his experiment to isolate certain types of neurons and compare their activity before and during and after learning. His experiment used rats' sense of smell in order to test neurons, as odor-related cognitive maps have been observed in rats before. The experimental setup is as follows.
Experimental Design Part 1
The rats are in a container with a head-sized hole and a feeder on either side. When the rat sticks its head into the hole, it receives an odor cue. These odors tell the rat where the reward will come from. There are 6 distinct odors, of order randomized between rats: two signal a reward from the right, two from the left, and two signal no reward. The rewards are split, with one odor cuing a grape flavor from the right, and another cuing a fruit punch flavor on the right. After training, the rats can reach 90% success rate with these cues.
Rationale
The experiment was designed this way to differentiate the types of neurons that are firing. There were four types of neurons that he considered.
- Reward Neurons
- These neurons fire differently for the reward vs no reward odors, but don't differentiate between right and left or flavors.
- Movement Neurons
- These neurons fire differently for right and left rewards, but don't care about flavor.
- Outcome Neurons
- These neurons fire differently for flavors
- Multiplex Neurons
- These neurons combine the three responses above in different ways.
Results Part 1
As stated, the rats were able to reach 90% accuracy in this experiment. They found that most neurons were multiplex; very few only served one purpose. In order to find which type of neuron was the most important in these responses, they trained a machine-learning model to predict the outcome from the neuron response. Using a confusion matrix, they found that the movement aspect of the experiment was the most important for the rats. The model confused the left-grape and left-fruit-punch responses with each other, but not with the right side. However, this was only true when they had already learned. He still needed to answer the question "How do they change while learning?" To test this, he designed a second phase of the experiment.
Experimental Design Part 2
For this stage of the experiment, they used 4 novel sets of odors for the same purpose. For each set of odors, they performed 5 training sessions with the rats. Each training session was composed of 400-500 data points.
Results Part 2
Over time, he found that the rat's mental representations of the experiment and its odors got simpler. Over the 5 sessions for a given odor set, they found that the rats improved their accuracy and narrowed their mental maps. The most important thing that they found was that the mental representations converged to a common mental image - that of the first phase of the experiment. Over time, he was able to predict the outcome of a test based on the neural maps from the original odor set! This is very cool.
Conclusion
In conclusion, this was a very very cool talk. He used tools that I had learned about before (voltage graphs, machine learning, confusion matrices) in a very direct way. I am very glad I went.