Embed Size (px)
Citation preview
nature neuroscience volume 12 | number 9 | SePTember 2009 1083
n e w s a n d v i e w s
7. Wigström, H. & Gustafsson, B. Nature 301, 603–604 (1983).
8. Carlson, G., Wang, Y. & Alger, B.E. Nat. Neurosci. 5, 723–724 (2002).
9. Klausberger, T. & Somogyi, P. Science 321, 53–57 (2008).
10. Freund, T.F. & Katona, I. Neuron 56, 33–42 (2007).11. Glickfeld, L.L. & Scanziani, M. Nat. Neurosci. 9,
807–815 (2006).12. Robbe, D. et al. Nat. Neurosci. 9, 1526–1533
(2006).13. Costa-Mattioli, M., Sossin, W.S., Klann, E. &
Sonenberg, N. Neuron 61, 10–26 (2009).14. Gundersen, V., Holten, A.T. & Storm-Mathisen, J. Mol.
Cell. Neurosci. 26, 156–165 (2004).15. Somogyi, J. et al. Eur. J. Neurosci. 19, 552–569
(2004).
involved in other notable effects of THC such as appetite stimulation and antiemesis.
1. Puighermanal, E. et al. Nat. Neurosci. 12, 1152–1158 (2009).
2. Katona, I. et al. J. Neurosci. 19, 4544–4558 (1999).
3. Kawamura, Y. et al. J. Neurosci. 26, 2991–3001 (2006).
4. Nyíri, G., Cserep, C., Szabadits, E., Mackie, K. & Freund, T.F. Neuroscience 136, 811–822 (2005).
5. Martin, B.R., Compton, D.R., Little, P.J., Martin, T.J. & Beardsley, P.M. NIDA Res. Monogr. 79, 108–122 (1987).
6. Monory, K. et al. PLoS Biol. 5, e269 (2007).
CB1 cannabinoid receptors on GABAergic axon terminals and suggest a mechanistic explanation for the cognitive impairment caused by acute THC administration at both the cellular and molecular level. A potential role for the mechanisms uncovered here in the chronic effects of cannabis consumption on memory and motivation is intriguing but remains to be clarified. Although this study focused on effects of THC in the hippocampus, it is likely that the mechanisms described here are pertinent to the effects of THC in other brain regions and may be
The author is in the Department of Physiology
and Biophysics and the Washington National
Primate Center, University of Washington School
of Medicine, Seattle, Washington, USA.
e-mail: [email protected]
Recognizing GrandmotherBharathi Jagadeesh
Using fMRI to probe face cells in the monkey temporal lobe, a study shows that these face-responsive cells appear to be feature detectors, but only work this way in the holistic construct of a face.
Humans, along with other primates, are very good at detecting and recognizing faces, sometimes finding them even when they are not there, as in the famous ‘face on Mars’ (Fig. 1a)1. This ability has long been thought to depend on the activity of temporal lobe cells that are selective for faces2–5. How do these cells produce selectivity for one class of images (faces) over other classes of images (objects)? Furthermore, how do these cells produce selectivity for the minor feature changes that define an individual face? Answering these questions is difficult for at least two reasons. First, we lack an appropriate, unbiased stimulus set of features that can be used to probe the selectivity of the wide variety of temporal lobe cells. Second, cells that are selective for a particular class of stimuli, such as faces, are encountered rarely5. Thus, selectivity is usually probed by making guesses about a potentially appropriate stimulus set that varies along a number of dimensions and by searching for cells modulated by this stimulus set. We therefore learn nothing more than the fact that many cells in the temporal lobe are selective for something and are modulated by the features that have been varied in a particular experiment, a circularity that is difficult to circumvent. The fundamental question
about whether the representation of faces is holistic (‘grandmother’ cells2 that respond to a specific face such as one’s grandmother; Fig. 1b) or made up of a combination of features3 has remained unanswered, in spite of years of study.
In the article by Friewald et al.6 in this issue, we receive a potential answer; at least one group of face cells appear to be feature detectors, but function as such only in the holistic construct of a face. The result is important, as the authors have been potentially successful in finding the right stimulus set with which to examine temporal lobe selectivity for faces. First, they used functional magnetic resonance imaging (fMRI) in monkeys to identify pockets of cells in the temporal lobe that are very likely to be selective for faces. By selecting a particular cluster of cells, the authors argue, they are likely to have limited the range of features for which a cell might be selective. The targeting limits the range of appropriate stimuli. The selective targeting also means that they can reliably find cells that are selective for faces, rather than relying on encountering rare face-selective cells5. Next, the authors simplified the potential space of face stimuli by using schematic ‘cartoons’ of faces (such as those shown in Fig. 1c), which contain a limited number of dimensions, to describe the complex physical characteristics of human faces.
Friewald et al.6 first verified that the cartoon faces produced responses similar to those produced by the photographs initially used to identify the patch (Fig. 1b,c); cartoon faces
produced nearly as large a response as the real faces. Next, the authors examined responses to parts of the cartoon faces (such as the eyes, nose and mouth), testing how well a cell responded to each part or combination of parts (Fig. 1d). All of the cells in the face patch were modulated by the presence of at least one, and at most four, parts of the face in the cell (Fig. 1d). Much of the variance in the response to the different face stimuli could be explained by the presence of one or two parts of the face. Face cells in the fMRI-identified face patch did not respond to the entire ‘grandmother’ (a full set of features and an outline corresponding to a specific face). Instead, they responded to one or two of her features, such as the eyes and mouth. Notably, the response to the features was interactive: the sum of the responses to all of the individual parts of a face was larger than the response to the entire face. In other words, ‘Grandmother’ produced a smaller response than the sum of her parts.
A particular grandmother has a unique pair of eyes, based on the different shapes the eyes can take. The authors tested the sensitivity of the cell to different dimensions of each feature. The seven parts that made up the cartoon faces in the study were varied along various dimensions (such as eye diameter; Fig. 1e). Almost all of the cells in the face patches were ‘tuned’ to the values of at least one feature dimension, with cells being tuned to about three features on average. Some dimensions were especially likely to modulate responses, particularly those expressed in the eyes and those that altered the overall face
©20
09 N
atu
re A
mer
ica,
Inc.
All
rig
hts
res
erve
d.
1084 volume 12 | number 9 | SePTember 2009 nature neuroscience
n e w s a n d v i e w s
cells, with nearby cells often representing different features in the face. Cells in the face patch thus respond to important features, but only in the context of a face. Therefore the answer to the question of whether face representation is holistic or based on the parts is that both are probably right.
How do these new results advance previous work? Previous work also suggests that eyes strongly modulate face cell responses5,7 and that a constellation of cells code a particular face, perhaps by representing different features in it5. However, these results are from a very small proportion of cells. In one study5, investigators had to record from over 800 cells before finding the 8% on which
sensitive (Fig. 1e) either in the presence of a mean features for all the other parts (Fig. 1e) or in the absence of the other parts of the face (Fig. 1e). The linear tuning for feature was preserved in the presence of average face features, but disappeared when the face context was removed; cells only signaled how big a grandmother’s eyes were if the eyes were placed in a face.
Cells tended to be tuned to a few features, resulting in an overlapping code among different neurons; thus, the presence of grandmother would result in a population response among neurons that responded to different characteristics in the face. However, there was no clear spatial organization of these
configuration. Many cells responded to changes in the dimension of the feature that they were selective for with a roughly linear change in response as a function of the dimension of the feature, producing the biggest responses to extreme versions of the feature. In other words, the bigger the grandmother’s eyes, the more these cells fired (Fig. 1e).
These experiments might suggest that these cells weren’t really ‘face’ cells at all, but that the cells that just represent the simpler shape contained in the face (the oblong shape of the eyes, for example). Did the response of these cells depend on the face at all? To answer this question, the authors tested the response to the one dimension that a cell seemed most
a
d e
b c
e
c
Eye size
Res
pons
e
2 features
1 feature
3 features
4 features Eyes scaled, other features random
Eyes scaled, other features at their means
Eyes scaled, other features removed
Grandmothers
Figure 1 Faces and features. (a) ‘Face on Mars’ taken in 1976 by the Viking 1 NASA spacecraft. (b) Faces of grandmothers showing variation in feature and identity. (c) Cartoon faces showing variation in feature and identity. (d) Cartoon faces consisting of 4, 3, 2 or 1 features. (e) Variation in eye size, from top to bottom, in the presence of other random face features, an average face feature or no face. Below, the response of cell is shown as a function of eye size in the presence of a variable face (solid line), a mean face (dashed line) or in the absence of a face (small dashed line).
Kim
Cae
sar
©20
09 N
atu
re A
mer
ica,
Inc.
All
rig
hts
res
erve
d.
nature neuroscience volume 12 | number 9 | SePTember 2009 1085
n e w s a n d v i e w s
these features be just as effective at driving cells in different face contexts, for example, in face composites? Will features identified in cartoon faces produce similar modulation in real faces? How do these cells respond to ambiguous9, imagined13 or learned5,7,9 faces? Will learning alter the feature dependence of responses8,11. It remains to be seen whether the coding described here still applies in these more complex situations.
1. Svoboda, E. Faces, faces, everywhere. New York Times, 13 February 2007.
2. Quiroga, R.Q., Reddy, L., Kreiman, G., Koch, C. & Fried, I. Nature 435, 1102–1107 (2005).
3. Desimone, R., Albright, T.D., Gross, C.G. & Bruce, C. J. Neurosci. 4, 2051–2062 (1984).
4. Yamane, S., Kaji, S. & Kawano, K. Exp. Brain Res. 73, 209–214 (1988).
5. Young, M.P. & Yamane, S. Science 256, 1327–1331 (1992).
6. Freiwald, W.A., Tsao, D.Y. & Livingstone, M.S. Nat. Neurosci. 12, 1187–1196 (2009).
7. Sigala, N. & Logothetis, N.K. Nature 415, 318–320 (2002).
8. Leopold, D.A., Bondar, I.V. & Giese, M.A. Nature 442, 572–575 (2006).
9. Tovee, M.J., Rolls, E.T. & Ramachandran, V.S. Neuroreport 7, 2757–2760 (1996).
10. Bar, M. et al. Proc. Natl. Acad. Sci. USA 103, 449–454 (2006).
11. Baker, C.I., Behrmann, M. & Olson, C.R. Nat. Neurosci. 5, 1210–1216 (2002).
12. Yamane, Y., Carlson, E.T., Bowman, K.C., Wang, Z. & Connor, C.E. Nat. Neurosci. 11, 1352–1360 (2008).
13. Kreiman, G., Koch, C. & Fried, I. Nature 408, 357–361 (2000).
modulates the responses of superior temporal sulcus face cells5. Another8 found that many face cells responded best to the average rather than to extreme values. These differences could potentially result from these studies not looking at the same temporal lobe face patch. This is an important caveat; the scheme here may not hold in other face patches. In other face-selective areas, learning5,7,9, adaptation or norming8, or hypothesis testing10 might be more important. This study6 offers hope for examining those cells as well, as fMRI identifies multiple face patches in the temporal lobe, potentially allowing each (and the interactions among them) to be examined in turn.
Cells in the face patches identified here are selective for a small number of features in a grandmother’s face, in the context of a larger face, and they are tuned to specific dimensions of this feature. These results vindicate some theories of visual object selectivity11. In addition, although the identification of a homogenous region of face cells in this study allowed for the testing of a large stimulus set, face space could still be more densely and systematically explored by recursive algorithms that explore relevant spaces12. Other questions also remain. Will
they could test tuning for different facial characteristics4,5,7. Face cells are rare, when one must stumble randomly through the temporal lobe looking for them.
This is not merely a technical weakness, as experiments almost always require searching for the relevant cells using the same stimulus set that will be ultimately used to test selectivity for features. Thus, the selection procedure can end up finding not the general properties of face cells, but face cells whose variance is likely to be explained by the feature variation contained in the stimulus set. Identifying the face patches independently, using faces in an independent fMRI experiment, as was done here, avoids this particular selection bias. Furthermore, the high likelihood of finding face selective cells in the patch means that tuning variations can be more systematically tested after finding the relevant cell. The greater yield (a technical advance) means that more hypotheses about tuning can be tested. Finally, because cells in the patch share characteristics, the likelihood that the sampled cells represent common properties of the local circuitry is increased.
At the same time, these results also conflict with other studies. For example, one study found that face familiarity significantly
Reconnecting injured nerves
Severed axonal connections do not spontaneously regenerate, creating a major hurdle for functional recovery following spinal cord injury (SCI). Previous attempts to aid axonal regeneration have failed to show correct reinnervation of specific target sites in the brain. On pp. 1106–1113 of this issue, Alto et al. demonstrate the successful anatomical regrowth of rat hind limb sensory nerve into the brainstem, across the SCI site.
The Tuszynski laboratory has previously shown that a combination of local neurotrophin expression, conditioning peripheral nerve injury, and cell grafts applied to and near a lesion site can provoke the partial regrowth of severed sensory axons after SCI. In these experiments, the lentiviral expression of neurotrophin-3 (NT-3) provided the chemoattractive guidance cue to direct regenerating axons. A bone marrow graft at the injury site provided scaffolding, a kind of cellular bridge, for the regenerating nerve tract. Sciatic nerve preconditioning lesions, a manipulation known to promote axonal regeneration, were also used. The regenerating axons grew into and beyond the spinal cord lesion site, but the regenerated axons in these previous studies did not quite reach their original targets in the brainstem.
Alto et al. now report ascending sensory tracts that successfully reconnect with their correct target site, the nucleus gracilis in the brain stem. To achieve this, the authors moved the injury site to level C1 on the spinal cord (compared with C4 in their earlier work), where the distance to the target site is much shorter (~2 mm). The picture shows a confocal micrograph of transganglionic labeling of injured sensory axons (cholera toxin B subunit labeling in red) reaching retrogradely labeled target neurons in the nucleus gracilis (Fluorogold labeling in green). To demonstrate the importance of local neurotrophin gradients in coaxing the regenerating axon to the correct site, the authors misexpressed NT-3 at an inappropriate location, the medullary reticular formation, and found regenerating axons being misdirected and inappropriately reconnected.
Disappointingly, despite this successful anatomical reconnection, these rats did not show appreciable functional recovery. Although regenerating axons formed ultrastructures that are consistent with de novo synaptic contacts, the target neurons in the brain stem showed little or no response to electrical stimulation of the sciatic nerve. Despite the lack of functional recovery, this study shows that proper chemotrophic guidance is a crucial step in promoting recovery after SCI. These findings can potentially aid in designing a combinatorial therapeutic strategy for individuals who have lost peripheral functions as a result of spinal cord trauma. Min Cho
©20
09 N
atu
re A
mer
ica,
Inc.
All
rig
hts
res
erve
d.
