Attention-like modulation of hippocampus place cell discharge

Abstract

Hippocampus place cell discharge is an important model system for understanding cognition, but evidence is missing that the place code is under the kind of dynamic attentional control characterized in primates as selective activation of one neural representation and suppression of another, competing representation. We investigated the apparent noise ("overdispersion") in the CA1 place code, hypothesizing that overdispersion results from discharge fluctuations as spatial attention alternates between distal cues and local/self-motion cues. The hypothesis predicts that: (1) preferential use of distal cues will decrease overdispersion; (2) global, attention-like states can be decoded from ensemble discharge such that both the discharge rates and the spatial firing patterns of individual cells will be distinct in the two states; (3) identifying attention-like states improves reconstructions of the rat's path from ensemble discharge. These predictions were confirmed, implying that a covert, dynamic attention-like process modulates discharge on a approximately 1 s time scale. We conclude the hippocampus place code is a dynamic representation of the spatial information in the immediate focus of attention.

Figure 1.

Overdispersion and the dynamic-attention hypothesis. A, Example of the unreliability of location-specific place cell discharge. The color-coded firing rate map for a place cell is shown (upper map) along with two 10 s paths that passed through the firing field (lower maps). The cell fired 66 action potentials (APs) along the pass depicted on the left and zero APs along the pass depicted on the right. Given the firing rate map, the likelihood of the two observations can be quantified by calculating a standardized firing rate for each pass (z). This calculation is depicted graphically using the histograms that correspond to each pass. On the assumption that the firing rate map describes the average rate in a location, integrating the rates along the path generates an expected number of APs during the pass. The expectation was 41.9 APs for the pass on the left and 10.5 APs for the pass on the right. The corresponding Poisson probability distribution of the number of expected APs for each pass is also shown. The standardized rate for a single pass is the normalized difference between the observed (obs) and expected (exp) activity, a standard normal deviate (z). Overdispersion is characterized by the variance of the distribution of standardized firing rates for a large number of passes. B, Diagram depiction of the dynamic-attention hypothesis that accounts for the place cell example in A. When the rat is at a location it chooses to attend to either distal cues (left) or local cues (right) to locate itself. The location-specific firing rate of each cell is modulated by the attentional state. In the raster of the diagramed cell, the cell fires more when the rat attends to distal cues (squares) and less when the rat attends to local cues (circles). According to this model, despite being at the same locations, the rat was in the attentional state associated with higher firing rate during the pass on the left and during the pass on the right the rat was in a different attentional state, one associated with lower firing rates. C, Example of how differences in place cell overdispersion will be compared. Two distributions of 1000 randomly-generated numbers with the averages equal to zero are depicted as histograms (left) and cumulative probability functions (right). The distributions only differ in that the variance of one distribution is 2 and the other is 4. It is easy to see that the distribution with the lower variance has the steeper cumulative probability function. The distributions intersect at probability = 50% because their medians are the same. The dynamic-attention hypothesis predicts that training rats to focus attention on one cue subset will reduce the variance of the standardized firing rates.

Figure 2.

Prediction 1. Navigation training reduced overdispersion and navigation training to use the distal cue subset reduced overdispersion even more. A, Schematic description of the behavioral conditions of the four groups of rats during the training and recording phases of the experiment. Group 1: Rats in the forage/stable group foraged for sporadically scattered food in a gray cylinder that was surrounded by curtains. A white card on the cylinder wall provided a polarizing landmark. The recording conditions were identical to the training conditions. Group 2: Rats in the forage/variable group foraged for scattered food in a black cylinder with a view of the surrounding room. The cylinder was stable on approximately half the training trials and it rotated at 1 rpm on the other trials. Only recordings during the stable trials were analyzed. Group 3: Rats in the navigate/stable group were also trained in the gray cylinder with a white card and surrounding curtains. The conditions were essentially the same as for the forage/stable rats except these rats performed a place preference task in which the food was only scattered whenever the rat visited an unmarked goal zone (white oval). Group 4: Rats in the navigate/variable group performed the same place preference navigation in the same physical conditions as the forage/variable group. On alternate trials the arena was either stable or rotating, but the navigation goal was always at a fixed location in the room. Although only recordings during the stable trials were analyzed, the rotating sessions discouraged the use of local cues for identifying the goal location because only distal cues could identify the stationary goal location. According to the dynamic attention hypothesis, overdispersion in the stable sessions should be less in the navigate/variable group than in the navigate/stable group even though the overt behavior was the same. This is because rats in the navigate/variable group were conditioned to attend to distal cues and to ignore local cues whereas rats in the navigate/stable group could always use both cue sets to locate the place preference goal. The environments also differed for the navigate/variable and navigate/stable groups, so the effect of the environment itself on overdispersion was evaluated by comparing overdispersion in the forage/stable and forage/variable groups. B, C, Overdispersion in the four groups was characterized as the variance (σ²) of the standardized firing rate along passes through a set of firing fields (n values given). Overdispersion was similar in the two forage groups and less in the navigate groups. D, Cumulative probability plots of the standardized firing rate distributions illustrate overdispersion was greatest in the forage groups, moderate in the navigate/stable group, and least in the navigate/variable group. E, Overdispersion was also characterized by analyzing the firing variance in the set of passes through an individual cell's firing field. Despite having different numbers of passes, each cell contributed one estimate of overdispersion. This analysis confirmed overdispersion was largest and similar in the two forage groups, moderate in the navigate/stable group and least in the navigate/variable group. (*p < 0.05).

Figure 3.

Differences on approaches to (red) and departures (blue) from the goal in the navigate/variable group and the relation of standardized firing to speed. A1–A4, The 5 s approaches to (left) and departures from (right) the goal in one session are superimposed on the gray-scaled session-averaged firing rate maps of 4 place cells. The gray-scale from light gray to black represents the range of firing rates from 0 AP/s to the peak (given below each map). White indicates the pixel was not visited. Gray-levels were assigned by linear interpolation of the rates between the centers of adjacent pixels in the averaged firing rate distributions. The track of the rat is superimposed on the rate map. Dots mark action potentials along the paths. Approaches are generally longer and straighter and cover more of the arena than departures. Nonetheless, firing is confined to a similar region on the approaches and departures. B, The speed on each approach and departure is plotted against the standardized firing rate for the behavioral episode. Although the speeds of approaches and departures overlap substantially, speed on approaches tend to be faster. The regression line (y = 0.15x − 0.32) on all the data indicates that speed is positively correlated with discharge but accounts for <1% of the variance (r = 0.09).

Figure 4.

Partitioning the local activity of a 53-place cell ensemble into two categories. A, Summary of the algorithm. Step 1: Generate an activity vector (AV) for each time step t_n. Step 2: Assign each AV to the average location in the time step. Step 3: Gather AVs for each location. Step 4. Partition the AVs at a location into two categories k1, and k2. B, Activity vectors partitioned in two ways for a 10 cm square pixel. B1, The vectors on the left were well-separated (separation quality = 0.56), close to the average yielded by the partitioning algorithm. The set of vectors on the right were randomly assigned to the two categories, yielding poor separation (separation quality = 0.01) that was typical for the random assignment. The random assignment was done so as to preserve the number of vectors in each category produced by the partitioning algorithm. The blue-to-red color code corresponds to 0 to 8 APs per 400 ms. B2, The well-separated and poorly-separated categorization was confirmed by principal component analysis. Plotting the first two principal components (PCA1, PCA2) for each AV illustrates that the AVs form two clusters after the partitioning but not after the random assignment. C, The AVs were well separated in far more pixels by the partitioning algorithm than by random assignment (*t₅₉ = 10.0; p ≈ 0).

Figure 5.

Distinct global ensemble activity states during foraging. A, Summary of the algorithm for assigning location-specific categories k1 and k2 across neighboring pixels to global states A and B. Step 1: activity vector correlations between category k1 and category k2 are computed for a base pixel and a neighboring pixel. The assignment to state A or B is determined as stated in the text. Step 2: regenerate the rat's trajectory tagging each data sample with one global state or the other. B, A correlation matrix comparing all pairs of 20 s ensemble activity vectors from a 10 min, 53-cell ensemble recording. The 20 s intervals are organized chronologically. C, The correlations of ensemble activity from moments in the same states (A-A, B-B) were greater than the correlations from moments when the ensemble was judged to be in different states (A-B, B-A). (*t₈₈₄ = 16.9; p ≈ 0).

Figure 6.

Prediction 2. Decoding of activity vectors produces distinct properties for individual place cells in each of the two global states. A, Firing rate maps from 7 cells of the 53-cell ensemble. For each cell the standard (raw) firing rate map is depicted along with the state-specific firing rate maps and the shuffled state maps, which help assess the likelihood that the decoding can occur by chance. The color code is the same for each set of 5 maps. The median firing rate in the red (peak rate) category is given below each map. The correlation between each shuffled pair or state-specific pair of firing rate arrays is given above the corresponding pairs of maps. The cells are sorted in ascending order of the state-specific correlation. B, Histograms of the signed rate distinction after random (top) and state-specific assignment (right). Note that the randomly generated histogram is unimodal and narrow whereas the histogram made with state-specific assignment is much broader and has a clear bimodal character. The implication is that there are approximately equal numbers of cells that discharge more rapidly in each of the states.

Figure 7.

Prediction 3. Decoding global ensemble states improves reconstruction of the rat's location. At each time step, an estimate of the decoded state of the current ensemble activity vector was either ignored (raw) or used to predict the current location. If the decoded state was used, the reconstruction was made by finding the best match between the vector and the state-specific template vectors that were constructed from the time series of state-specific activity vectors. The state-specific template vectors were either always chosen to match the state of the current ensemble vector (decoded), or be opposite to that state (opposite). A, Three 10 s paths that were reconstructed at 400 ms and 2.5 cm resolution. The rat's path (black) and the reconstructed path using the decoded ensemble state (red) were typically more similar than the estimates from the raw data (white) or estimates based on the incorrect ensemble state (blue). The numbers give the average prediction error in pixels for the path. B, Summary of 16 reconstruction attempts with the 53-cell ensemble. Current position was reconstructed at 400, 800, 1200, and 1600 ms temporal resolutions and at 2.5, 5, 10, and 20 cm spatial resolutions. The average error from each of the 16 combinations of the temporal and spatial resolutions is represented as a color-coded matrix. The maximum error in each color category is given; hotter colors indicate better accuracy. Reconstruction accuracy using the decoded spike trains appears equivalent or better for each attempt. C, Summary of 48 reconstruction attempts. Location was reconstructed for three recordings (ensemble sizes = 54, 53, and 23 place cells) at the four temporal and spatial resolutions. The average error for each set of reconstructions was normalized by the average reconstruction error using the observed, state-independent data: $\frac{\left(\text{state − observed}\right)}{\left(\text{state + observed}\right)}$. Reconstructions using the decoded state (red) were always better than reconstructions from the raw state-independent data. Not only was this improvement significant (t₄₇ = 9.4; p ≈ 0), but it was a result of accurate state decoding, because reconstructions using the incorrect state (blue) were often worse than the raw, state-independent predictions (paired t₄₇ = 2.2; p = 0.03).