The Behavioral Relevance of Cortical Neural Ensemble Responses Emerges Suddenly

Abstract

Whereas many laboratory-studied decisions involve a highly trained animal identifying an ambiguous stimulus, many naturalistic decisions do not. Consumption decisions, for instance, involve determining whether to eject or consume an already identified stimulus in the mouth and are decisions that can be made without training. By standard analyses, rodent cortical single-neuron taste responses come to predict such consumption decisions across the 500 ms preceding the consumption or rejection itself; decision-related firing emerges well after stimulus identification. Analyzing single-trial ensemble activity using hidden Markov models, we show these decision-related cortical responses to be part of a reliable sequence of states (each defined by the firing rates within the ensemble) separated by brief state-to-state transitions, the latencies of which vary widely between trials. When we aligned data to the onset of the (late-appearing) state that dominates during the time period in which single-neuron firing is correlated to taste palatability, the apparent ramp in stimulus-aligned choice-related firing was shown to be a much more precipitous coherent jump. This jump in choice-related firing resembled a step function more than it did the output of a standard (ramping) decision-making model, and provided a robust prediction of decision latency in single trials. Together, these results demonstrate that activity related to naturalistic consumption decisions emerges nearly instantaneously in cortical ensembles. Significance statement: This paper provides a description of how the brain makes evaluative decisions. The majority of work on the neurobiology of decision making deals with "what is it?" decisions; out of this work has emerged a model whereby neurons accumulate information about the stimulus in the form of slowly increasing firing rates and reach a decision when those firing rates reach a threshold. Here, we study a different kind of more naturalistic decision--a decision to evaluate "what shall I do with it?" after the identity of a taste in the mouth has been identified--and show that this decision is not made through the gradual increasing of stimulus-related firing, but rather that this decision appears to be made in a sudden moment of "insight."

Keywords: attractor; cortex; decision; gustatory; hidden Markov models; taste.

Figure 1.

Late-onset cortical activity reflects behavioral responses to the stimuli. A, The amount that rats choose to consume (in terms of mean number of licks assessed in a brief access task; Breslin et al., 1993) of each stimulus varies as a function of both the chemical identity and concentration of that stimulus. This schematic distribution serves as the basis for assessing the correlation with behavioral (consumption-related) choice. B, The trial-averaged, stimulus-aligned responses of three cortical neurons. Typical of such responses, initial activity was not stimulus specific, selectivity emerging only after ∼200 ms (Katz et al., 2001; Piette et al., 2012; Sadacca et al., 2012). Still later, for each neuron, the response became obviously choice related, with the strongest responses for this neuron to the most aversive stimuli and the weakest to the most palatable stimuli. C, The apparent emergence of choice-related responses in B is confirmed using a moving-window analysis of the linear correlation between the numbers of spikes in B and the behavior pattern in A.

Figure 2.

Taste identification precedes taste evaluation. The ramp of choice-related activity evident in single neurons (Fig. 1C) is consistent across the entire neural sample (black line; ±2 SEM; n = 68), though lower than for individual exemplars because of the inclusion of nonresponsive neurons in the analysis. This palatability-related firing ramped upward, achieving significance above baseline at 0.83 s after stimulus (p < 0.05, Tukey–Kramer test). The onset of evaluative coding occurred only long after tastes could be identified by ensemble activity, as individual taste pairs (gray dashed lines) were reliably discriminated above chance by 0.4 s following taste delivery, and 100% classification was achieved across all taste pairs (solid gray line) by 0.6 s after stimulus. The duration of each curve, as determined by the time period spanning between 30 and 70% of the total ramp height, for taste identification and palatability coding are displayed (horizontal lines) above the individual curves.

Figure 3.

Ensemble cortical responses form reliable sequences of states with coherent, trial-specific state-to-state transition times. A, Each neuron's spiking probability (per 10 ms) is plotted for each of five states (color coded to B) that occurred in the hidden Markov model solution derived for sucrose. B, HMM-determined probability that a set of simultaneously recorded cortical neurons achieves each firing-rate state (colored curves), plotted together with ensemble spiking activity (each vertical notch represents a spike), for four consecutive trials of one stimulus (sucrose). The same sequence was identified in most trials (here, the first 3 of 4 trials), but the times of state-to-state transitions varied from trial to trial [periods of high state likelihood (>80%) are highlighted in color]. C, The time courses of the HMM-derived late state (i.e., the state dominant after 1 s) probability for all sucrose trials in one session, revealing both the reliability and suddenness of this state's emergence (it progresses from 0 probability to 1.0 probability across a <100 ms period in almost every trial) and the considerable variability in the state's onset latency from trial to trial. The solid black line shows the time-average probability across trials, which forms a gradual ramp with a time course reminiscent of Figure 2. D, The distribution of identified late state onsets across all modeled ensembles (42 models total).

Figure 4.

Realignment of cortical ensemble data to the appropriate HM state onsets sharpens the emergence of choice-related firing. A, The emergence of choice-related firing in ensemble activity is sharper following realignment of each trial's spiking activity to the onset of the state identified as dominant at 1 s after stimulus delivery (PTTH; red dashed line) than when that activity is aligned to stimulus delivery (PSTH; black dashed line). Sigmoidal fits to each time series are overlain (solid lines) and provide estimates for the speed with which activity transitions to choice relatedness. B, The values and 95% confidence intervals for the duration of the transition from low levels of choice-related firing to asymptotic choice-related firing (the time across which the slope of the fit sigmoid curve is highest, parameter 1/β) for PSTH and PTTH data, in addition to two control realignments: data realigned to either the state prior to the identified late state (prelate aligned) or the state dominant during taste identification (early aligned). In all cases, real data transitioned more quickly than the PSTH data (Z_PSTH = 10.6, p_PSTH > 0.001) or either of the control alignments (Z_EARLY = 6.4, p_EARLY > 0.001, Z_PRELATE = 3.5, p_PRELATE > 0.001).

Figure 5.

The sharpening of choice-related firing by alignment to HM states is significantly greater than expected by chance. A, Single-trial raster plots (top) and peristimulus spiking probabilities (bottom) for two example cortical single-neuron taste responses (in this case, to quinine; red trace) compared with a simulation of spiking generated from the mean response of that neuron (gray trace). B, To estimate whether the sharpening shown in Figure 4 was merely an effect of HMM realignment, regardless of genuine rapid transitions in the neural data, trials from the experimental data were reshuffled (PSTH shuffled; lighter traces) and were modeled using HMM (PTTH shuffled; darker maroon traces). Realignment produced a modest sharpening in this control data. For a second control data set (as plotted in A), spike trains were simulated from the PSTHs of real neurons, maintaining the average activity of the ensemble, minus moment-to-moment correlations among neurons (PSTH simulation; light gray traces). These simulated data were also modeled using HMM (PTTH simulation; dark gray traces). Again, realignment caused a modest sharpening of choice-related activity. C, The results of sharpening 100 simulated data sets and 100 shuffled data sets are compared to the effect of sharpening the experimental data. Only 4 of the 200 simulations and 7 of the 200 shuffled data sets transitioned into choice-related firing as suddenly as the real data, confirming that the effect shown in A and B was not merely an effect of HMM-cued realignment (p_simulated < 0.05; p_shuffled < 0.05; one-tailed binomial test).

Figure 6.

The emergence of choice-related firing in cortical ensembles resembles instantaneous state transitions. A, Single-trial raster plots (top) and peristimulus spiking probabilities (bottom) for two example cortical single-neuron taste responses (to quinine). Shown are the corresponding simulated neural spike trains and mean responses generated from a ramping DDM (green traces) and from an instantaneous rate-state transition model (OPTI; blue traces) for these two neurons. B, The simulated data were modeled in the same manner as the real ensembles and PSTH-derived simulations (Fig. 3), and choice related activity was calculated before and after HMM state alignment. Emergence of choice-related firing in both DDM- and OPTI-derived simulations (light shading, stimulus aligned; dark shading, state aligned) substantially benefited from realignment. C, When directly compared to the suddenness of choice-related activity onset in real cortical ensembles, the instantaneously transitioning OPTI data (blue bars) both transition significantly more rapidly than this DDM (green bars) simulation (p < 0.01, Kolmogorov–Smirnov test) and are a better match to the real data (Λ_OPTI-DDM = 41.8, p < 0.001).

Figure 7.

The onset of the palatability-related state predicts consumption decisions. A, Four consecutive quinine trials (time after stimulus delivery on the x-axis), showing the HMM solution (state probability on the y-axis). The state represented by a dashed red line is the palatability-related state. Overlain on this presentation are the times at which the rat gaped (vertical hash marks). The time of the first gape is the behavioral readout of the decision to reject the quinine stimulus. B, For every trial in the session from which the panels in A were culled, the latencies of the palatability-related state (x-axis) are plotted against the latencies of gaping onset (y-axis). The thick diagonal dashed line is unity: above this line, the ensemble transition preceded the making of the decision; below, the opposite is true. In the vast majority of trials, the ensemble transition preceded the behavioral choice with a latency of <500 ms (thinner diagonal dashed line). C, Across eight sessions (see Materials and Methods), the middle 50% of times to first gape were closer to the median (p < 0.03) when timing was aligned to the HM transition (red) than when timing was aligned to the stimulus onset (dark gray) or to trial-shuffled transitions (light gray). The panel shows times normalized to entire distribution width. The inset shows the absolute time.