Concept Cells through Associative Learning of High-Level Representations
2014; Cell Press; Volume: 84; Issue: 2 Linguagem: Inglês
10.1016/j.neuron.2014.10.004
ISSN1097-4199
Autores Tópico(s)Zebrafish Biomedical Research Applications
ResumoIn this issue of Neuron, Quian Quiroga et al., 2014Quian Quiroga R. Kraskov A. Mormann F. Fried I. Koch C. Neuron. 2014; 84 (this issue): 363-369Abstract Full Text Full Text PDF Scopus (27) Google Scholar show that neurons in the human medial temporal lobe (MTL) follow subjects' perceptual states rather than the features of the visual input. Patients with MTL damage however have intact perceptual abilities but suffer instead from extreme forgetfulness. Thus, the reported MTL neurons could create new memories of the current perceptual state. In this issue of Neuron, Quian Quiroga et al., 2014Quian Quiroga R. Kraskov A. Mormann F. Fried I. Koch C. Neuron. 2014; 84 (this issue): 363-369Abstract Full Text Full Text PDF Scopus (27) Google Scholar show that neurons in the human medial temporal lobe (MTL) follow subjects' perceptual states rather than the features of the visual input. Patients with MTL damage however have intact perceptual abilities but suffer instead from extreme forgetfulness. Thus, the reported MTL neurons could create new memories of the current perceptual state. Neurons along the ventral visual pathway respond with varying degrees of specificity to subjects' perceptual decisions. In situations where the visual input and the subjective percept can be experimentally dissociated, most neurons in early visual areas respond to low-level stimulus properties, whereas approximately 90% of neurons in higher-level inferotemporal (IT) cortex are modulated by the subjects' perceptual report (Logothetis, 1998Logothetis N.K. Philos. Trans. R. Soc. Lond. B Biol. Sci. 1998; 353: 1801-1818Crossref PubMed Scopus (304) Google Scholar). Neurons from area TE of IT feed into medial temporal lobe (MTL) structures that include the hippocampal formation and the entorhinal, perirhinal, and parahippocampal cortices (Suzuki and Eichenbaum, 2000Suzuki W.A. Eichenbaum H. Ann. N Y Acad. Sci. 2000; 911: 175-191Crossref PubMed Scopus (134) Google Scholar). A new study in this issue of Neuron by Quian Quiroga et al., 2014Quian Quiroga R. Kraskov A. Mormann F. Fried I. Koch C. Neuron. 2014; 84 (this issue): 363-369Abstract Full Text Full Text PDF Scopus (27) Google Scholar shows that "concept cells" in the human MTL closely follow subjective awareness. Several studies have demonstrated the presence of "concept cells" in the human MTL (Quian Quiroga, 2012Quian Quiroga R.Q. Nat. Rev. Neurosci. 2012; 13: 587-597PubMed Google Scholar). Concept cells are highly selective neurons that seem to represent the meaning of a given stimulus in a manner that is invariant to different representations of that stimulus. For example, a single neuron in the human hippocampus was found to selectively respond to several different pictures of the actress Halle Berry, even when she was disguised as Catwoman, the role she played in one of her movies. The same neuron also responded to the letter string "HALLE BERRY" but not to other letter strings. Later studies showed that these "concept cells" were also activated when stimulus information was provided in other sensory modalities, for example, hearing the name of a person spoken aloud (Quian Quiroga, 2012Quian Quiroga R.Q. Nat. Rev. Neurosci. 2012; 13: 587-597PubMed Google Scholar). These invariant, multimodal responses are in contrast to the more stimulus-specific responses observed in visual area IT and suggest that MTL neurons encode stimulus information in an abstract form, such that various instances of a given stimulus can activate these neurons. In the new study, Quian Quiroga et al., 2014Quian Quiroga R. Kraskov A. Mormann F. Fried I. Koch C. Neuron. 2014; 84 (this issue): 363-369Abstract Full Text Full Text PDF Scopus (27) Google Scholar ask whether concept cells encode the perceptual reports of subjects when visual inputs are ambiguous. Human subjects were presented with ambiguous images created by morphing two different pictures (for example, a morph between Presidents Bill Clinton and George Bush). Before viewing the morphed image, subjects viewed an adaptor image (a picture of Clinton or of Bush) and then reported with whom the morphed image corresponded. As expected from previous adaptation studies, exposure to the adaptor image biased subjects' perception, with the result that the subsequent morphed image was identified as the opposite of the adaptor. Thus, in this paradigm, the very same morphed image could be identified either as Clinton or Bush depending on which adaptor stimulus was used. By recording from a neuron selective to Clinton or Bush, Quian Quiroga et al., 2014Quian Quiroga R. Kraskov A. Mormann F. Fried I. Koch C. Neuron. 2014; 84 (this issue): 363-369Abstract Full Text Full Text PDF Scopus (27) Google Scholar could then examine how the neuron's response varied with the perceptual report of the subject. Quian Quiroga et al., 2014Quian Quiroga R. Kraskov A. Mormann F. Fried I. Koch C. Neuron. 2014; 84 (this issue): 363-369Abstract Full Text Full Text PDF Scopus (27) Google Scholar recorded from 62 neurons in different MTL structures and, consistent with the evidence for high-level representations in MTL neurons, found that rather than merely signaling the visual input, the activity of MTL neurons correlated with the subjects' perceptual decision. Strikingly, in many cases, the responses to the morphed images were indistinguishable from the responses to the original nonmorphed image, both when considering the magnitude as well as the latency of responses (although Quian Quiroga et al., 2014Quian Quiroga R. Kraskov A. Mormann F. Fried I. Koch C. Neuron. 2014; 84 (this issue): 363-369Abstract Full Text Full Text PDF Scopus (27) Google Scholar are careful to point out that this null result could be due to the large variability in responses across neurons). That is, in our example of the Clinton neuron above, on trials when the morphed images were identified as Clinton, the Clinton neuron's response was no different from its responses to the nonmorphed image of Clinton, regardless of the degree of morphing. Thus, these neurons appear to respond in an all-or-none manner, their response reflecting the conscious percept of the subject, regardless of the ambiguity in the visual input. These findings along with results from studies employing the mental imagery (Kreiman, 2007Kreiman G. Curr. Opin. Neurobiol. 2007; 17: 471-475Crossref PubMed Scopus (19) Google Scholar), binocular rivalry (Kreiman, 2007Kreiman G. Curr. Opin. Neurobiol. 2007; 17: 471-475Crossref PubMed Scopus (19) Google Scholar), change detection (Reddy et al., 2006Reddy L. Quiroga R.Q. Wilken P. Koch C. Fried I. Curr. Biol. 2006; 16: 2066-2072Abstract Full Text Full Text PDF PubMed Scopus (36) Google Scholar), and masking (Quian Quiroga, 2012Quian Quiroga R.Q. Nat. Rev. Neurosci. 2012; 13: 587-597PubMed Google Scholar) paradigms reveal that the activity of MTL neurons correlates strongly with conscious vision, rather than the visual features of the input stimuli. What could be the functional role of these neurons? On the face of it, these neurons have properties that almost make them look like potential candidates for being the elusive neural correlate of consciousness (NCC) (Crick and Koch, 1990Crick F. Koch C. Cold Spring Harb. Symp. Quant. Biol. 1990; 55: 953-962Crossref PubMed Scopus (217) Google Scholar). However, this seems implausible since patients with bilateral MTL damage show no obvious problems in perceptual awareness and consciousness (Squire et al., 2004Squire L.R. Stark C.E. Clark R.E. Annu. Rev. Neurosci. 2004; 27: 279-306Crossref PubMed Scopus (2015) Google Scholar). This suggests that the NCC lies upstream of these MTL neurons. Instead, patients with MTL damage exhibit profound memory impairments and MTL structures have long been known to play a key role in recognition and associative memory processes (Squire et al., 2004Squire L.R. Stark C.E. Clark R.E. Annu. Rev. Neurosci. 2004; 27: 279-306Crossref PubMed Scopus (2015) Google Scholar, Suzuki and Eichenbaum, 2000Suzuki W.A. Eichenbaum H. Ann. N Y Acad. Sci. 2000; 911: 175-191Crossref PubMed Scopus (134) Google Scholar). Single-neuron studies in the nonhuman primate have revealed that MTL neurons signal the learning of associations or relationships between stimulus pairs. A preferred cue as well as its learned paired associate activated pair-coding neurons in perirhinal cortex (Naya et al., 2003Naya Y. Yoshida M. Miyashita Y. J. Neurosci. 2003; 23: 2861-2871PubMed Google Scholar), and the learning of arbitrary associations between stimuli and spatial locations has been observed in hippocampal neurons (Wirth et al., 2003Wirth S. Yanike M. Frank L.M. Smith A.C. Brown E.N. Suzuki W.A. Science. 2003; 300: 1578-1581Crossref PubMed Scopus (288) Google Scholar). fMRI studies have also shown that the MTL encodes associations between stimuli even when subjects are unaware of the temporal contingencies between them (Schapiro et al., 2012Schapiro A.C. Kustner L.V. Turk-Browne N.B. Curr. Biol. 2012; 22: 1622-1627Abstract Full Text Full Text PDF PubMed Scopus (238) Google Scholar). This ability of MTL neurons to encode associations between initially unrelated stimuli presented in close temporal proximity could underlie the formation of concept cells. Indeed, the statistics of the world ensure that different representations of a given stimulus (e.g., a person's face and hearing her name pronounced) co-occur frequently. A simple associative learning mechanism in conjunction with the statistical regularities of the world could result in MTL neurons encoding different forms of a given stimulus, thus exhibiting the property of invariance, a key feature of concept cells (Figure 1). The new study by Quian Quiroga et al., 2014Quian Quiroga R. Kraskov A. Mormann F. Fried I. Koch C. Neuron. 2014; 84 (this issue): 363-369Abstract Full Text Full Text PDF Scopus (27) Google Scholar unequivocally demonstrates that MTL concept cells follow the visual percept of the subject. However, as we have argued above, MTL neurons are unlikely to be the direct correlates of subjective awareness. Instead, it is probable that, in cooperation with neocortex, these neurons play a key role in transforming the current perceptual state of the subject into long-term memories. As mentioned above, the activity of most neurons in IT cortex correlates with subjective visual awareness. This information about the perceptual state of the subject enters the MTL in the perirhinal cortex (Suzuki and Eichenbaum, 2000Suzuki W.A. Eichenbaum H. Ann. N Y Acad. Sci. 2000; 911: 175-191Crossref PubMed Scopus (134) Google Scholar). Patients with MTL damage have difficulty creating new memories, and lesion studies have shown that connections from MTL structures to neocortex are necessary for establishing and maintaining long-term memories in area TE (Higuchi and Miyashita, 1996Higuchi S. Miyashita Y. Proc. Natl. Acad. Sci. USA. 1996; 93: 739-743Crossref PubMed Scopus (196) Google Scholar). Thus, it is plausible that the functional role of concept cells is to create new memories of the current perceptual experience of the subject based on information they receive about high-level representations in cortex. In addition, the abstract nature of concept cell responses implies that the new memories could be linked to other memories of the same stimulus, such that future viewings of the stimulus could potentially activate all memories related to this stimulus. One intriguing possibility is that MTL neurons select patterns of activity in neocortical structures that are sufficiently long lasting to merit long-term storage in episodic memory. Studies using rapid sequential visual presentation (RSVP) show that high-level processing is possible even when the input image changes 75 times a second (Potter et al., 2014Potter M.C. Wyble B. Hagmann C.E. McCourt E.S. Atten. Percept. Psychophys. 2014; 76: 270-279Crossref PubMed Scopus (239) Google Scholar). Single-unit recording studies have also demonstrated that even at such high presentation rates, IT neurons can show a brief "blip" of activity each time the cell's preferred stimulus is presented (Keysers et al., 2001Keysers C. Xiao D.K. Földiák P. Perrett D.I. J. Cogn. Neurosci. 2001; 13: 90-101Crossref PubMed Scopus (227) Google Scholar). But not all such briefly presented stimuli are remembered. Perhaps, MTL circuits are designed to only store patterns that remain active in cortex for a certain minimum duration, for instance, at least 100–200 ms (Thorpe, 2012Thorpe S.J. Grandmother cells and Distributed Representations.in: Kriegeskorte N. Kreiman G. Visual Population Codes: Toward a Common Multivariate Framework for Cell Recording and Functional Imaging. MIT Press, Cambridge2012: 23-52Google Scholar). The hippocampus could thus act as a gatekeeper, letting information through into episodic memory only if the corresponding cortical representations persist beyond this temporal threshold. This process of pattern selection could explain why MTL neurons have such long latencies relative to neurons in the neocortex. An elegant way of implementing such a pattern selection mechanism could be to use an oscillatory buffer in which any neural representation that is maintained on two consecutive oscillatory cycles would be selected for memory storage. Such a (wildly speculative) mechanism could rely on the prominent theta oscillations observed in MTL structures. In conclusion, evidence is accumulating that "concept cells" carry high-level, abstract stimulus information. The new study by Quian Quiroga et al., 2014Quian Quiroga R. Kraskov A. Mormann F. Fried I. Koch C. Neuron. 2014; 84 (this issue): 363-369Abstract Full Text Full Text PDF Scopus (27) Google Scholar nicely underscores this idea by demonstrating that these cells track the high-level perceptual report of subjects, even when the incoming stimulus information is noisy or ambiguous. As we have argued above, these MTL cells are most likely postperceptual, not required for supporting perception, but for encoding the current percepts into new episodic memories. Single-Cell Responses to Face Adaptation in the Human Medial Temporal LobeQuian Quiroga et al.NeuronSeptember 25, 2014In BriefUsing an adaptation paradigm to bias the perception of morphed faces, Quian Quiroga et al. show how single neurons in the human medial temporal lobe follow the subjective perception by the subjects rather than the visual features of the stimulus. Full-Text PDF Open Access
Referência(s)