Fast readout of object identity from macaque inferior temporal cortex

Science

Volumn 310, Issue 5749, 2005, Pages 863-866

  Hung, Chou P ^a,b,c   Kreiman, Gabriel ^a,b,c   Poggio, Tomaso ^a,b,c   DiCarlo, James J ^a,b,c  

^a MASSACHUSETTS INSTITUTE OF TECHNOLOGY   (United States)

^b MASSACHUSETTS INSTITUTE OF TECHNOLOGY   (United States)

^c MASSACHUSETTS INSTITUTE OF TECHNOLOGY   (United States)

Author keywords

[No Author keywords available]

Indexed keywords

BIOTECHNOLOGY; COMPUTATION THEORY; INFORMATION TECHNOLOGY; NEUROLOGY; OBJECT RECOGNITION;

BRAIN COMPUTATION; INFERIO TEMPPRAL (IT) CORTEX; OBJECT IDENTITY; POPULATION LEVEL;

BRAIN;

BIOLOGY;

ACCURACY; ANIMAL CELL; ANIMAL EXPERIMENT; ARTICLE; BRAIN; CELL POPULATION; CONTROLLED STUDY; INFERIOR TEMPORAL CORTEX; INFORMATION PROCESSING; MACACA; NERVE CELL; NONHUMAN; POSITION; PRIORITY JOURNAL; RECOGNITION; TEMPORAL CORTEX;

ACTION POTENTIALS; ANIMALS; BRAIN MAPPING; MACACA MULATTA; NEURONS; PSYCHOMOTOR PERFORMANCE; RECOGNITION (PSYCHOLOGY); TEMPORAL LOBE; TIME FACTORS; VISUAL PERCEPTION;

ANIMALIA; MACACA;

EID: 27644475783     PISSN: 00368075     EISSN: 10959203     Source Type: Journal    
DOI: 10.1126/science.1117593     Document Type: Article

References (42)

1. M. Fabre-Thorpe, G. Richard, S. J. Thorpe, Neuroreport 9, 303 (1998).
2. D. Perrett, J. Hietanen, M. Oeam, P. Benson, Philos. Trans. R. Soc. 355, 23 (1992).
3. N. K. Logothetis, D. L. Sheinberg, Annu. Rev. Neurosci. 19, 577 (1996).
4. K. Tanaka, Annu. Rev. Neurosci. 19, 109 (1996).
5. K. Grill-Spector, R. Malach, Annu. Rev. Neurosci. 27, 649 (2004).
6. C. Keysers, D. K. Xiao, P. Foldiak, D. I. Perret, J. Cogn. Neurosci. 13, 90 (2001).
7. R. vanRullen, S. Thorpe, Vision Res. 42, 2593 (2002).
8. E. Rolls, Curr. Opin. Neurobiol. 1, 274 (1991).
9. R. Desimone, T. Albright, C. Gross, C. Bruce, J. Neurosci. 4, 2051 (1984).
10. P. M. Gochin, M. Colombo, G. A. Dorfman, G. L. Gerstein, C. G. Gross, J. Neurophysiol. 71, 2325 (1994).
11. B. Richmond, R. Wurtz, T. Sato, J. Neurophysiol. 50, 1415 (1983).
12. N. K. Logothetis, J. Pauls, T. Poggio, Curr. Biol. 5, 552 (1995).
13. L. F. Abbott, E. T. Rolls, M. J. Tovee, Cereb. Cortex 6, 498 (1996).
14. S. L. Brincat, C. E. Connor, Nat. Neurosci. 7, 880 (2004).
15. M. Ito, H. Tamura, I. Fujita, K. Tanaka, J. Neurophysiol. 73, 218 (1995).
16. G. Wallis, E. T. Rolls, Prog. Neurobiol. 51, 167 (1997).
17. M. Riesenhuber, T. Poggio, Nat. Neurosci. 2, 1019 (1999).
18. D. Freedman, M. Riesenhuber, T. Poggio, E. Miller, Science 291, 312 (2001).
19. H. Tomita, M. Ohbayashi, K. Nakahara, I. Hasegawa, Y. Miyashita, Nature 401, 699 (1999).
20. A. Ishai, L. G. Ungerleider, A. Martin, J. L. Schouten, J. V. Haxby, Proc. Natl. Acad. Sci. U.S.A. 96, 9379 (1999).
21. T. Poggio, S. Smale, Notices Am. Math. Soc. 50, 537 (2003).
22. K. V. Shenoy et al., Neuroreport 14, 591 (2003).
23. L. F. Abbott, Q. Rev. Biophys. 27, 291 (1994).
24. R. deCharms, A. Zador, Annu. Rev. Neurosci. 23, 613 (2000).
25. A. S. Tolias, A. G. Siapas, S. M. Smimakis, N. K. Logothetis, Soc. Neurosci. Abstr. 557.5 (2002).
26. A. Pouget, P. Dayan, R. Zemel, Annu. Rev. Neurosci. 26, 381 (2003).
27. C. Koch, Biophysics of Computation (Oxford Univ. Press, New York, 1999).
28. http://cbcl.mit.edu/software-datasets/kreiman/ fastreadout.html.
29. The eight groups used for categorization [toys (t), foods (fd), human faces (hf), monkey faces (mf), hand/body parts (h), vehicles (v), box outlines (b), cats/dogs (cd)] were defined before the experiments. Unsupervised clustering of neuronal responses yielded similar groups (28). Categorization became substantially worse upon arbitrarily defining these groups as sets of random objects (28). The discriminability for individual sites for passive and active viewing were similar (fig. S7, Supporting Online Material).
30. We assumed independence among neurons; this assumption should be revisited upon recording simultaneously from many neurons because correlations may contain additional information. Our estimate represents a lower bound on the information represented by small neuronal populations. However, even under these conditions, we obtain a high degree of accuracy [see also (41)].
31. Throughout the paper we randomly selected a given number of sites for decoding. The brain could be selectively wired such that targets of IT receive stronger input from the most relevant features. A simple feature selection step before the input to the classifier to select the sites with the highest signal-to-noise ratio (Supporting Online Material) showed that high performance levels could be achieved using a much smaller number of sites (fig. 52).
32. R. Quiroga, L. Reddy, G. Kreiman, C. Koch, I. Fried, Nature 435, 1102 (2005).
33. M. P. Young, S. Yamane, Science 256, 1327 (1992).
34. i) had very similar performance and were no better than the regularized linear classifiers for n > 64 sites. The estimated coefficients depend on regularization and are different for different regularization techniques (27).
35. Multiple sources of noise can affect the encoding of information. The performance of the classifier was very robust to deletions of substantial numbers of neurons during testing, simulating neuronal or synaptic death (fig. S1A), and also to large proportions of deleted spikes (simulating failures in spike transmission or neurotransmitter release; fig. S1B).
36. We trained the classifier for the categorization task with 70% of the pictures and then tested it on the remaining 30% of the pictures. The performance was quite good and only slightly below the performance levels reported above (fig. S3; compare to Fig. 1).
37. Y. Sugase, S. Yamane, S. Ueno, K. Kawano, Nature 400, 869 (1999).
38. R. Vogels, Eur. J. Neurosci. 11, 1239 (1999).
39. B. Tjan, Adv. Neural Inf. Processing Syst. 13, 66 (2001).
40. T. Poggio, E. Bizzi, Nature 431, 768 (2004).
41. N. C. Aggelopoulos, L. Franco, E. T. Rolls, J. Neurophysiol. 93, 1342 (2005).
42. We thank M. Kouh for the object recognition model; R. Quiroga and A. Kraskov for spike sorting; J. Mayo and J. Deutsch for technical support; CSBI for computer cluster usage; and R. Desimone, N. Kanwisher, C. Koch, and T. Serre for comments on the manuscript This research was sponsored by grants from NIH. NSF, and especially from the Defense Advanced Research Projects Agency and Office of Naval Research. Additional support was provided by Eastman Kodak Company, Daimler Chrysler, Honda Research Institute, The Pew Charitable Trusts, Whiteman fellowship (C.K.), and the McDermott chair (T.P.).