New in Neuroscience: Do neurons gradually ramp or instantly step to a decision?

We make decisions from the moment we get up (what to eat for breakfast? what should I wear today?) until the moment we go to sleep (do I actually need to brush my teeth?). Decisions are important – some of them guide the course of our entire lives. How does our brain carry out this complex and highly important process? To answer this question, neuroscientists study the activity of decision-related neurons – cells whose activity strongly correlate with our choices. One question that has been asked by decision-making researchers for over 20 years is whether these neurons gradually change their activity as a decision is being made, indicating that a single neuron gradually accumulates evidence for or against a decision, or if these neurons fully ‘commit to a decision’ at once with a discrete jump in activity. For the most part, scientists assumed the former, but Latimer et al. in fact show the latter – suggesting that while people make decisions gradually, neurons don’t!

The ramping model (left) vs. the stepping model (right).

To investigate this question, Latimer et al. analyzed the activity of neurons in the lateral intraparietal (LIP) cortex, a region known to contain these decision-related neurons, while a monkey performed a decision-making task. This task is relatively simple: the animal must determine which direction dots move on a screen. However, these dots are shown along with noise, akin to ‘snow’ on a TV screen, making the task more difficult and forcing the monkey to take some time in determining average dot motion direction. While the monkey is deciding, neurons in LIP will increase their activity if the monkey thinks motion is towards one direction and decrease if motion appears to be in the opposite direction. Typically, scientists look at the average activity over many trials of this task, and observe a gradual increase or decrease - hence the 20-year assumption of gradual change in activity. But, there’s a problem with this: taking the average over many trials may smear out what’s really happening on a single trial.

While it may seem obvious that the next step should be to simply analyze activity separately for each trial, it’s actually difficult to do this and pin down whether neurons gradually or instantly increase activity. Why? Because when neurons are ‘active’, they emit electrical pulses called spikes (higher activity means more spikes in a shorter time), and these spikes are a little bit random. These noisy spikes make it difficult to figure out if activity looks more ramp-like or step-like because for most trials, neural activity looks like it could be a noisy version of either. To get around this problem and answer the question at hand, the authors formulated two models of neural spiking: a ramping model in which activity gradually increases or decreases, and a stepping model in which activity instantly jumps up or down (once). They then asked, which model was more likely to generate the type of spiking activity observed in the monkey? To answer this, they simulated these models to generate ‘fake’ trains of spikes, and compared the similarity of model activity to the real data. They found that the stepping model does a better job of matching the data at both the single-trial and averaged-trial level, suggesting that neurons flip from ‘undecided’ (medium activity) to ‘decided’ (high or low activity) at a single point in time.

While the study was convincing, there are still a couple things to check. For example – what if the monkey decided quickly, say in 100 ms, but the authors considered activity for the entire 1000 ms trial? Would analyzing only the first 100 ms for these trials change the result? While the authors did preliminary work for these concerns, more work and replication of this result will be required to know for sure. The results of this study also raise many more questions: do all neurons in LIP step together? Are there other places in the brain that exhibit ramping? Future experiments will certainly answer these questions, but for the moment, this work serves as a nice first ‘step’ in understanding how we implement this vitally important process.