In this study, we found that single neurons of the human hippocampus exhibit two distinct memory signals. One signal is generic: The recorded neurons respond differentially depending on the categorical status of the test item (novel vs. repeated). This signal was found in all four brain regions that we examined (hippocampus, amygdala, prefrontal cortex, and anterior cingulate). The other memory signal is item-specific: Each neuron responded strongly to a small fraction of repeated words, and each repeated word elicited strong responding in a small fraction of neurons. This signal was found only in the hippocampus.
The method for identifying the item-specific signal in the hippocampus involved comparing the shapes of neuron-by-item normalized spiking distributions for novel vs. repeated items. This method of analysis is not intuitive and likely would not have been pursued absent predictions made by neurocomputational models (16-19). Those models predict that pattern-separated, item-specific memory representations are sparsely coded, in which case the memory signals they generate should be hard to detect. The models further predict that, for test items equated in every respect except their episodic occurrence in the experimental context (novel vs. repeated), these signals should be selectively detected for repeated items. The QQ plots of data from the hippocampus confirmed this prediction (Fig. 2A). This elusive memory signal was predicted to exist only in the hippocampus and, indeed, was observed only in the hippocampus.
Previous single-unit studies with humans typically searched only for the generic memory signal, finding it in the amygdala and the hippocampus (13-15), and later in parietal cortex (22). Similarly, we found the generic memory signal in all four brain regions that were examined. Thus, this signal is not only generic (responding to the categorical status of test items), it is also widespread.
What role might the widespread generic memory signal play? Hippocampal neurons that sparsely code item-specific, episodic memories may distribute generic information to other brain regions (and to other neurons in the hippocampus) to perform memory-related functions not requiring item-specific information. For example, the prefrontal cortex may determine whether the memory signal associated with a test item is strong enough to exceed a decision criterion. Because that decision would depend on memory strength, represented by the generic signal, item-specific content would not be required. Similarly, parietal cortex may play a role in assessing confidence (23, 24), which again requires only information about the strength of the memory signal. On this view, the item-specific memory signal in the hippocampus is fundamental, whereas the generic memory signal – the focus of much prior research – is secondary and derivative.
Our findings contradict a recent claim (25) that episodic memories in the human hippocampus are not stored as sparsely-coded, pattern-separated representations. The proposal is that concept neurons code episodic memories. Concept neurons are neurons that fire selectively when specific concepts (e.g., “Eiffel Tower”) are evoked, whether the evoking stimulus is a picture representing the concept, its printed name, or its spoken name (26). Critically, concept neurons can expand their tuning, responding to unrelated stimuli that have been paired with the relevant concept (27). For example, if the Eiffel Tower and the actor Jackie Chan are shown together during an experiment, the “Eiffel Tower neuron” will subsequently increase its firing rate to a Jackie Chan image presented alone.
In our view, these neural signals do not reflect episodic memories. Concept neurons respond to any stimulus that evokes their preferred concepts (e.g., the Eiffel Tower), which suggests that they code factual knowledge (i.e., semantic memory). Neurons that code episodic memory code the episodic aspects of a memory, namely, what, when, and where (3). As noted by Suthana et al. (28), the tasks used in concept cell studies do not require the patient to appreciate what, when, and where information about test stimuli. By contrast, our study required the patient to appreciate precisely that kind of information, and it yielded clear evidence of sparsely-coded, pattern-separated episodic memory signals in the hippocampus. Studies using other methodologies have also identified pattern-separated memory signals in the hippocampus (29-31)
We measured differences in single-unit activity for novel vs. repeated words: Each word was presented twice differing only in its status as being novel or repeated. Concept neurons that respond to semantic meaning would respond to both the first and second presentation of a word. By contrast, we described neurons that fired differentially for particular repeated words and that therefore integrated stimulus information about the attributes of “what” (i.e., this particular word), “when” (presented a few minutes ago), and “where” (in this experimental context). This item-specific, neural code was sparse, pattern-separated, and limited to the hippocampus.