Single Neuron Coding of Identity in the Human Hippocampal Formation

Abstract

Experimental findings show the ubiquitous presence of graded responses and tuning curves in the neocortex, particularly in visual areas [1-15]. Among these, inferotemporal-cortex (IT) neurons respond to complex visual stimuli, but differences in the neurons' responses can be used to distinguish the stimuli eliciting the responses [8, 9, 16-18]. The IT projects directly to the medial temporal lobe (MTL) [19], where neurons respond selectively to different pictures of specific persons and even to their written and spoken names [20-22]. However, it is not clear whether this is done through a graded coding, as in the neocortex, or a truly invariant code, in which the response-eliciting stimuli cannot be distinguished from each other. To address this issue, we recorded single neurons during the repeated presentation of different stimuli (pictures and written and spoken names) corresponding to the same persons. Using statistical tests and a decoding approach, we found that only in a minority of cases can the different pictures of a given person be distinguished from the neurons' responses and that in a larger proportion of cases, the responses to the pictures were different to the ones to the written and spoken names. We argue that MTL neurons tend to lack a representation of sensory features (particularly within a sensory modality), which can be advantageous for the memory function attributed to this area [23-25], and that a full representation of memories is given by a combination of mostly invariant coding in the MTL with a representation of sensory features in the neocortex.

Keywords: hippocampus; invariance; medial temporal lobe; memory; single neurons.

Figure 1

Exemplary Invariant Units (A) Responses of a unit in the left hippocampus. For each stimulus, the raster plot (blue lines represent the appearance of a spike and each row is associated to a trial; first trial is at the top and time zero is the stimulus onset) and instantaneous firing rate are shown. Stimulus numbers appear next to the stimulus pictures. The unit responded to actor “Jackie Chan” but not to another actor, “Luciano Castro.” (B) Responses of a unit in the left hippocampus. The unit responded to all stimuli associated to a TV host (“Topa”) but not to pictures of a cat or the written or spoken name “Gato” (cat). For space reasons, only 10 of 20 stimuli are shown on each panel, but there were no significant responses to the other stimuli not shown. See Figure S1 for further examples of response-eliciting stimuli and Figure S2 for another example of a neuron responding to multiple identities.

Figure 2

Response Characteristics of the Population of Recorded Cells (A) Normalized firing rate for all response-eliciting stimuli, separated by stimulus type (picture, text, and sound). Within the pictures, the 47 response-eliciting identities are separated by black lines and sorted by their shortest latency. Response latency for each stimulus is denoted by a red star. Figure S2 shows a neuron with 15 response-eliciting stimuli from three different identities. (B) The proportion of response-eliciting identities was evaluated in different conditions and segregated between the units in the hippocampus and amygdala. In both regions, similar proportions were found for visual stimuli (for visual invariance and for the proportion of identities showing responses to more than one picture), but larger proportions in the hippocampus were seen for text and sound stimuli. In the case of triple invariance, i.e., identities showing visual invariance and responses to the text and sound stimuli, the proportion of identities in the hippocampus was significantly larger than in amygdala (Z test, p < 0.05). (C) Histograms for the normalized strength of activity in all the responsive units, computed for all the stimuli presented in each session and separated according to whether or not they were responsive. (D) Strength (baseline corrected) and latency for each picture in 25 out of the 47 response-eliciting identities (the remaining 22 are presented in Figures S3A and S3B). Each triplet associated to a given identity is depicted as a triangle of a given color with each picture corresponding to the vertices of the triangle. (E) Average joint distance (in the strength-latency space presented in Figure 3D) for the three pictures of a given identity (left), and triplets created by randomly selecting three pictures out of the 141 pictures (right). Mean ± standard error of the mean is shown in red, while boxplots are in blue (center line, median; box limits, upper and lower quartiles; notch limits, (1.57 × interquartile range)/sqrt(n)) with the whiskers in black extending to the most extreme data points not considered as outliers. There were significant differences between the two populations (rank-sum test, p ∼10⁻¹⁷). See Figures S3E and S3F for the independent analysis of strength and latency.

Figure 3

Comparison of Responses within and between Identities (A) Percentage of pairs exhibiting significant differences in strength (left) and latency (right) according to a permutation test for the response-eliciting pairs “within identity.” The pairs could be formed by two responsive pictures (PIC versus PIC), one responsive picture and the written name from the same identity (PIC versus TXT), or one responsive picture and the spoken name from the same identity (PIC versus SND). The percentage was significantly smaller for the PIC versus PIC case when compared with the others. ^∗p < 0.05; ^∗∗p < 10⁻⁶. An alternative analysis was performed using a decoding approach. Figure S4 shows an example of a decoder being unable to significantly distinguish among the stimuli within a given identity. (B) Proportion of pairs exhibiting significant differences in strength (left) and latency (right) according to a permutation test for the response-eliciting pairs “within identity” (from the current dataset) and “between identities” (from the dataset reported in [26], see STAR Methods). These proportions showed no significant differences across the datasets for strength (Z test, p = 0.15), but the proportion for the within identity dataset was significantly smaller than for the between identities dataset (Z test, p = 7.3 × 10⁻³). (C) Distributions of the normalized difference in strength of response for each dataset. Vertical arrows denote the median of the corresponding distributions. There was no significant difference between them (rank-sum test, p = 0.12). (D) Same as in (C) but for differences in latency. We observed that the latency difference was significantly smaller for within identity pairs than for between identities pairs (rank-sum test, p ∼10⁻⁸).

Figure 4

Schematic Summarizing Some of the Results When a picture from Bill Clinton is shown, a cell assembly of MTL neurons will fire in response to the stimulus. When another picture from Bill Clinton is presented, the set of MTL neurons firing has a large overlap (∼80%) with the previous assembly. When the written name is presented, the overlap of activated neurons is much lower (∼40%), as the very different sensory information carried by the name leads to very distinct cortical inputs to the MTL. In turn, the results from [28] showed that when the picture of Hillary Clinton is presented, a highly associated concept to Bill Clinton, a small overlap of ∼4% will be activated by both stimuli. Still, this small proportion is much larger than the < 1% observed for non-associated pictures (e.g., Bill Clinton and Lionel Messi), providing a substrate for encoding a meaningful association.