Direct observation of the neural computations underlying a single decision

  1. Department of Neuroscience, Columbia University, New York, NY 10027, USA
  2. Zuckerman Mind Brain Behavior Institute, Columbia University, New York, NY 10027, USA
  3. McGovern Institute for Brain Research, Massachusetts Institute of Technology, Cambridge, MA 02139, USA
  4. Grossman Center for the Statistics of Mind, Columbia University, New York, NY 10027, USA
  5. Howard Hughes Medical Institute, Columbia University, New York, NY 10027, USA
  6. Kavli Institute for Brain Science, Columbia University, New York, NY 10027, USA


  • Reviewing Editor
    Tobias Donner
    University Medical Center Hamburg-Eppendorf, Hamburg, Germany
  • Senior Editor
    Michael Frank
    Brown University, Providence, United States of America

Reviewer #1 (Public Review):


In this paper, Steinemann et al. characterized the nature of stochastic signals underlying the trial-averaged responses observed in the lateral intraparietal cortex (LIP) of non-human primates (NHPs), while these performed the widely used random dot direction discrimination task. Ramp-up dynamics in the trial averaged LIP responses were reported in numerous papers before. However, the temporal dynamics of these signals at the single-trial level have been subject to debate. Using large-scale neuronal recordings with Neuropixels in NHPs, allows the authors to settle this debate rather compellingly. They show that drift-diffusion-like computations account well for the observed dynamics in LIP.


This work uses innovative technical approaches (Neuropixel recordings in behaving macaque monkeys). The authors tackle a vexing question that requires measurements of simultaneous neuronal population activity and hence leverage this advanced recording technique in a convincing way.

They use different population decoding strategies to help interpret the results.

They also compare how decoders relying on the data-driven approach using dimensionality reduction of the full neural population space compare to decoders relying on more traditional ways to categorize neurons that are based on hypotheses about their function. Intriguingly, although the functionally identified neurons are a modest fraction of the population, decoders that only rely on this fraction achieve comparable decoding performance to those relying on the full population. Moreover, decoding weights for the full population did not allow the authors to reliably identify the functionally identified subpopulation.


No major weaknesses.

Reviewer #2 (Public Review):

Steinemann, Stine, and their co-authors studied the noisy accumulation of sensory evidence during perceptual decision-making using Neuropixels recordings in awake, behaving monkeys. Previous work has largely focused on describing the neural underpinnings through which sensory evidence accumulates to inform decisions, a process which on average resembles the systematic drift of a scalar decision variable toward an evidence threshold. The additional order of magnitude in recording throughput permitted by the methodology adopted in this work offers two opportunities to extend this understanding. First, larger-scale recordings allow for the study of relationships between the population activity state and behavior without averaging across trials. The authors' observation here of covariation between the trial-to-trial fluctuations of activity and behavior (choice, reaction time) constitutes interesting new evidence for the claim that neural populations in LIP encode the behaviorally-relevant internal decision variable. Second, using Neuropixels allows the authors to sample LIP neurons with more diverse response properties (e.g. spatial RF location, motion direction selectivity), making the important question of how decision-related computations are structured in LIP amenable to study. For these reasons, the dataset collected in this study is unique and potentially quite valuable.

However, the analyses at present do not convincingly support two of the manuscript's key claims: (1) that "sophisticated analyses of the full neuronal state space" and "a simple average of Tconin neurons' yield roughly equivalent representations of the decision variable; and (2) that direction-selective units in LIP provide the samples of instantaneous evidence that these Tconin neurons integrate. Supporting claim (1) would require results from sophisticated population analyses leveraging the full neuronal state space; however, the current analyses instead focus almost exclusively on 1D projections of the data. Supporting claim (2) convincingly would require larger samples of units overlapping the motion stimulus, as well as additional control analyses.

Specific shortcomings are addressed in further detail below:

  1. The key analysis-correlation between trial-by-trial activity fluctuations and behavior, presented in Figure 5-is opaque, and would be more convincing with negative controls:

To strengthen the claim that the relationship between fluctuations in (a projection of) activity and fluctuations in behavior is significant/meaningful, some evidence should be brought that this relationship is specific - e.g. do all projections of activity give rise to this relationship (or not), or what level of leverage is achieved with respect to choice/RT when the trial-by-trial correspondence with activity is broken by shuffling.

  1. The choice to perform most analysis on 1D projections of population activity is not wholly appropriate for this unique type of dataset, limiting the novelty of the findings, and the interpretation of similarity between results across choices of projection appears circular:

The bulk of the analyses (Figure 2, Figure 3, part of Figure 4, Figure 5, Figure 6) operate on one of several 1D projections of simultaneously-recorded activity. Unless the embedding dimension of these datasets really does not exceed 1 (dimensionality using e.g. participation ratio in each session is not quantified), it is likely that these projections elide meaningful features of LIP population activity. Further, additional evidence/analysis would help to strengthen the authors' interpretation of the observed similarity of results across these 1D projections. For one, the rationale behind deriving Sramp was based on the ramping historically observed in Tin neurons during this task, so should be expected to resemble Tin. Second, although Tin does not comprise the majority of neurons recorded in each session, it does comprise the largest fraction of the neuron groups (e.g. Tin, Min, etc) sampled during most sessions, so SPC1 should be expected to resemble Tin more than it does the other neuron groups. Additional/control analyses will be important for strongly supporting the claim that the approximate equality between the population DV and the average of Tin units is meaningful. The analysis presented in Figure S7 is an important step toward this, demonstrating that SPC1 isn't just reflecting the activity of Tin, but would make the point more strongly with some additional analysis. Are the magnitudes of weights assigned to units in Tin larger than in the other groups of units with pre-selected response properties? What is their mean weighting magnitude, in comparison with the mean weight magnitude assigned to other groups? What is the null level of correspondence observed between weight magnitude and assignment to Tin (e.g. a negative control, where the identities of units are scrambled)?

A secondary approach could also get at this point (the small Tin group furnishes a DV very similar to the overall population DV) from a different direction: computing SPC1 using only neurons *not* in Tin, and repeating the analysis performed with the other 3 1D projections of the data currently in Figure 5. Observing similar results for this 4th projection would strengthen the evidence supporting the interpretation the authors adopt.

  1. The principal components analysis normalization procedure is unclear, and potentially incorrect and misleading:

Why use the chosen normalization window (+/- 25ms around 100ms after motion stimulus onset) for standardizing activity for PCA, rather than the typical choice of mean/standard deviation of activity in the full data window? This choice would specifically squash responses for units with a strong visual response, which distorts the covariance matrix, and thus the principal components that result. This kind of departure from the standard procedure should be clearly justified: what do the principal components look like when a standard procedure is used, and why was this insufficient/incorrect/unsuitable for this setting?

  1. Analysis conclusions would generally be stronger with estimates of variability and control analyses: This applies broadly to Figures 2-6.

Reviewer #3 (Public Review):


The paper investigates which aspects of neural activity in LIP of the macaque give rise to individual decisions (specificity of choice and reaction times) in single trials, by recording simultaneously from hundreds of neurons. Using a variety of dimensionality reduction and decoding techniques, they demonstrate that a population-based drift-diffusion signal, which relies on a small subset of neurons that overlap choice targets, is responsible for the choice and reaction time variability. Analysis of direction-selective neurons in LIP and their correlation with decision-related neurons (T con in neurons ) suggests that evidence integration occurs within area LIP.


This is an important and interesting paper, which resolves conflicting hypotheses regarding the mechanisms that underlie decision-making in single trials. This is made possible by exploiting novel technology (Primatepixels recordings), in conjunction with state-of-the-art analyses and well-established dynamic random dot motion discrimination tasks.

  1. Howard Hughes Medical Institute
  2. Wellcome Trust
  3. Max-Planck-Gesellschaft
  4. Knut and Alice Wallenberg Foundation