Article Figures & Data
Figures
- Figure 1.
Schematic of spiking properties of recorded SC neurons. A, A widely accepted model of the left SC topographic map and the corresponding right visual hemifield. When a visual stimulus is presented in a particular location and/or a saccade is made to that location, neurons are active across the SC map. The hot spot of activity for the example recorded location is at the green dot when the amplitude and direction of the stimulus/saccade are 20° and 0°, respectively; activity spreads spatially across the SC in a Gaussian-like manner. B, C, SC activity elicited for vectors 45° (B) and −45° (C) away from the preferred direction of the recorded neuron. These hypothetical cases highlight how two very different direction vectors can elicit similar activity levels at the recorded location. Figure panels A through C adapted from Gandhi and Katnani (2011). The same conceptual quandary remains even if the topographic map is updated to reflect unequal representations of upper and lower hemifields (Hafed and Chen, 2016). D, Traditional single electrode approach into the SC (left) compared with an advanced recording technique with a multichannel laminar probe (right). In both cases, the insertion angle is orthogonal to the SC surface, yielding neuron(s) at only one location on the SC map (e.g., the location of the green dot in the previous panels). Figure panel adapted from Jagadisan and Gandhi (2022).
- Figure 2.
Peri-event values of spiking activity and LFPs simultaneously recorded across channels. A, The across-trials mean firing rates for all 15 functional channels recorded during an example session are plotted aligned to target onset (left) and saccade onset (right) to eight radially equidistant targets. Each colored trace represents the spiking activity on one channel averaged across all trials to a particular target. Subplots are rotated so that the preferred target direction of this population is displayed horizontal and rightward with respect to center. B, The across-trials mean LFP voltage values for all 15 channels are plotted using the same conventions as the spiking activity data.
- Figure 3.
Linear discriminant classification of spiking and LFP activity. Sliding 100-ms windows of summed spike counts or average LFP voltage on each channel were used to train a linear discriminant analysis (LDA) model and test its ability to decode target direction. Mean (±SEM) across-session classifier performances for the spike count (black traces) and LFP classifiers (green traces) are plotted separately for each of eight target directions and aligned to target onset (left panels) or saccade onset (right panels). Chance level classifier performance was obtained by using shuffled class labels during the training phase. Performance values were grouped across sessions by aligning to each session’s preferred target direction (visualized here as the right middle panels), and the performance for each session and each target was baseline-subtracted before averaging. Values for each window are plotted aligned to the end of that window (e.g., performance of the classifier trained and tested on the 0- to 100-ms window following target onset is plotted at 100 ms on the x-axis).
- Figure 4.
Spread of spatial direction discrimination across broad visual space. Summary polar plots of mean across-session classifier performance distribution across target directions during each epoch as defined in Materials and Methods for spiking (A) and LFP (B) activity. Spatial tuning of spiking activity is broader in the motor epoch than any other epoch. For LFPs, decoding performance is lower during the delay period but is comparable between the visual and motor epochs.
- Figure 5.
Linear discriminant classification of spiking and LFP activity: systematic variation of bin width. A, Classifiers were trained and tested on summed spike counts during windows of lengths ranging from 20 to 300 ms. Average performance values over 50 bootstrapping iterations for each target direction and each window length condition is plotted using the same conventions as Figure 3 for one example session. Again, values are plotted aligned to the end of each window; therefore, each condition peaks in classification performance at different times, but this is not the comparison of interest. Spike count-based classification is largely robust to window size during the transient visual and motor epochs (as indicated by the dark blue and light green arrows at Target 1) but performance increases with increasing window sizes during the delay period. B, As in A but for average LFP voltage on each channel during windows of varying lengths. A decrease in performance with increasing window lengths can be seen during the motor epoch (indicated by dark blue and light green arrows at Target 1), but the opposite effect can be seen during the delay period.
- Figure 6.
Linear discriminant classification of spiking and LFP activity: systematic variation of population size. A, Classifiers were trained and tested on summed spike counts during 100-ms windows with randomly selected population sizes ranging from 1 to 17 channels. Average performance over 50 bootstrapping iterations for each target direction and each population size condition are plotted using the same conventions as Figure 3 for one example session. As population size increases, classification performance increases in a corresponding fashion. B, As in A but for classifiers based on average LFP voltage across a varied number of included channels (matched to the channels included in the spike count classifiers). LFP-based classifier performance also increases systematically as a function of population size.
- Figure 7.
Classification performance as a function of population size during four key epochs. The performance of spike count (black traces) and LFP (green traces) classifiers was evaluated through a systematic variation of population size (Fig. 6). Here, the across-session average peak classification performance for each target during the visual (blue panels), early delay (light purple), late delay (dark purple), and motor (orange panels) epochs is plotted as a function of the number of channels included (from 1 to U; see Materials and Methods). During both the visual and motor epochs, increasing population size leads to a corresponding increase in direction discriminability, even for targets in the hemifield opposite the preferred direction. For spike count-based classifiers, performance in the delay period follows the same trend, whereas less consistency is observed in the performance of LFP-based classifiers during these epochs.
- Figure 8.
Comparison of direction encoding during the visual and motor epochs for each target. Peak decoding performance in the visual (x-axis) versus motor (y-axis) epoch as defined in Materials and Methods for each target. Spike-based classifiers are indicated in black and LFP-based classifiers are indicated in green. Each session (N = 18) contributes two points to each of the eight target subplots: one for spiking activity and another for LFP activity. Inset, Significant (paired t test) differences in performance level during the visual and motor epochs for each target are represented, with p < 0.05 indicated by a single asterisk, p < 0.01 by double asterisks, and p < 0.001 by triple asterisks. For spike-based classifiers, the performance is significantly different between epochs for all targets but one. For LFP-based classifiers, only targets far from the preferred direction have significantly different encoding across epochs.
- Figure 9.
Comparison of spatial encoding properties of spiking and LFP activity across epochs. A, Baseline-shifted classification performance on spiking (black) and LFP (green) activity during each of the four main epochs (as defined in Materials and Methods) for each target aligned to the preferred direction of the population. Mean across sessions (bold lines) as well as each session’s individual tuning curve (N = 18, thin lines) are shown. Session-averaged traces are the same as the data shown in Figure 4. B, Differences in the amount of spatial information encoded between two signal modalities. Trapezoidal area under each observed tuning curve (AUC) shown in A was computed, and the LFP classifiers’ AUCs were subtracted from the spike count classifiers’ AUCs in a pairwise fashion for each session and epoch. The across-session mean difference in AUC between the two modalities (bold line) and individual session values (gray points) are plotted. Significant differences between spiking and LFP classifier distributions are shown with asterisks at the α = 0.05 significance level (paired t test; p < 0.05 is indicated by a single, p < 0.01 double, and p < 0.001 triple asterisk). From the visual epoch, the encoding of spatial information is significantly different between spiking and LFP signals. C, The AUC during each epoch for each session (thin lines) along with the across-session mean AUC (bold lines) were computed after shifting each population’s decoding values such that the decoding performance was 1 for the target in the preferred direction (i.e., Target 1). This measure allows for a fair comparison of breadth of information across epochs. D, Grid of statistical differences (paired t test) in tuning width across pairs of epochs computed separately for each signal modality. For spiking activity, the tuning width is only significantly different between the visual and motor epochs and between the late delay and motor epochs. For LFPs, the tuning width is significantly different across all epochs.
In this issue
