Abstract
Simultaneous recording from many neurons in macaque lateral intraparietal area reveals the elusive drift-diffusion signal, long suspected to underlie individual perceptual decisions and response times.
Introduction
For decades, the "drift-diffusion model" (DDM) has been the gold standard for explaining perceptual decision-making. The theory suggests that the brain accumulates noisy evidence over time until it hits a threshold, triggering a choice. However, because older technology only allowed researchers to record one or two neurons at a time, they had to average data across thousands of trials to see a clear signal.
This "trial-averaging" reveals the drift (the average rate of evidence accumulation) but completely hides the diffusion (the trial-by-trial noise and variability). This left a fundamental gap: Does the DDM actually describe the neural computations of a single decision, or is it just a mathematical abstraction that happens to work on average? Steinemann et al. aimed to bridge this gap by observing the decision process unfold in a single brain, on a single trial, in real-time.
Methods
The researchers recorded from the lateral intraparietal area (LIP) of macaque monkeys using Neuropixels 1.0-NHP probes. These high-density probes allowed them to monitor up to 203 neurons simultaneously across eight recording sessions, yielding 1,084 total neurons.
To extract the "decision signal" from this high-dimensional population activity, they employed several dimensionality reduction and decoding strategies:
- Targeted Regression: Identifying a "coding direction" in the neuronal state space that best approximated a linear ramp on each trial.
- Principal Component Analysis (PCA): A hypothesis-free method to see if the dominant variance in the population corresponded to the decision.
- What/When Decoders: Training logistic decoders to predict what choice the monkey would make and when it would initiate the eye movement.
- Mediation Analysis: A statistical test showing that the neural signal at an early time point influenced the final choice specifically by affecting the state of the neural signal at a later time point—proving a continuous accumulation process.
Key Insights and Results
The "Aha" of Single Trials
Unlike trial-averaged ramps, single-trial signals are meandering and non-monotonic. They look like "biased Brownian motion," exactly as predicted by the diffusion component of the DDM. The variance of these single-trial signals increased linearly over time, a classic signature of a diffusion process.
Dimensionality and Specialization
Despite the complexity of the brain, the information used to make these choices is remarkably low-dimensional. A single scalar quantity derived from the population activity was sufficient to predict both choice and reaction time. PCA revealed that the first three principal components explained ~67% of the task-related variance, with PC1 alone capturing the decision-related "ramp."
Furthermore, the researchers identified two distinct "gears" in the LIP machinery:
- The Integrators ($T_{in}$): A small subset of neurons (about 15%) whose response fields overlap the choice targets. These neurons represent the integral (the running sum) of evidence.
- The Integrands ($M_{in}$): Direction-selective neurons that represent the integrand (the raw, noisy evidence being fed into the sum).
Discussion and Limitations
While this study provides a "smoking gun" for the DDM in LIP, it remains correlational. The authors acknowledge that the decision variable is likely distributed across multiple brain areas, such as the PFC and FEF. Additionally, the subset of neurons representing the "integrand" was small, limiting some statistical analyses regarding evidence routing.
Future work may involve simultaneous recordings from the sensory source (Area MT) and the integrator (LIP) to directly observe the "handshake" between sensory evidence and cognitive accumulation. This research marks a significant shift from trial-averaging toward understanding the real-time computational dynamics of the individual brain.
Summary updated