視覚5
Vision 5
O3-7-1-1
色の集団符号化におけるカテゴリカルな変調
Categorical modulation in population coding of color

○田嶋達裕1, 鯉田孝和2, 田嶋千尋3, 合原一幸3, 鈴木秀幸3, 小松英彦4
○Satohiro Tajima1, Kowa Koida2, Chihiro I. Tajima3, Kazuyuki Aihara3, Hideyuki Suzuki3, Hidehiko Komatsu4
NHK技研1, 豊橋技科大エレクトロニクス先端融合研2, 東大生研3, 生理研4
NHK STRL, Tokyo1, EIIRIS, Toyohashi Univ of Tech, Toyohashi2, IIS, Univ of Tokyo, Tokyo3, NIPS, Okazaki4

Categorical structure of recognition is known to have profound effects on perception and neural representation of stimulus, in a wide variety of sensory modalities and features. In this study, we analyzed responses of color-selective neurons recoded in the macaque IT cortex, which change their activities depending on the task demands: discrimination or categorization [1]. To clarify the functional meanings of the task-specific modulation, we focused on the stimulus representation at the level of neural population rather than at single neurons. We found that the task demands modulated the entire shape of activity distribution over the neural population. The analysis of temporal structure revealed that the modulation component had a stimulus-dependent dynamics after the stimulus presentation, suggesting that the categorical modulation can be driven according to the visual input. The polarity of modulation was consistent to the categorical biasing effect, which is predicted by a recurrent network model that approximates the optimal probabilistic inference of stimulus dynamics based on a hidden-Markov model [2]. (References: [1] Koida, K., Komatsu, H. (2007) Nat. Neurosci., 10(1):108-116. [2] Imai, C., Tajima, S., Aihara, K., Suzuki, H. (2011). Proc. 16th Artificial Life and Robotics, 350-353.)
O3-7-1-2
マカク属サル下側頭葉の局所細胞集団によって表出される特徴の数について
The number of independent features represented by local cell population in the macaque inferotemporal cortex

○横尾征一郎1, 最上嗣生1, 田中啓治1
○Seiichirou Yokoo1, Tsuguo Mogami1, Keiji Tanaka1
理化学研究所 脳科学総合研究センター 認知機能表現研究チーム1
Lab. for Cognitive Brain Mapping, RIKEN Brain Science Institute1

Stimulus selectivity of single cells in macaque inferotemporal (IT) cortex is determined in the domain of moderately complex features, and cells responding to similar features cluster in a local columnar region of IT. However, little is known about relation between object categories and IT columns. To determine how cell population in a column is organized in relation to object categories, we recorded many (30~65) cells from each column and examined their responses to 850 object images (50 members each of 17 categories). About 90% of cells recorded in a column were category selective, and 30~90% of cell pairs had similar category selectivity: mean responses to 17 categories were positively correlated between the two cells. For the cell pairs with similar category selectivity, the similarity in stimulus selectivity was examined in the correlation between responses to 50 individual stimuli belonging to a commonly effective category. Whereas 30~80% of cell pairs showed significantly positive correlation, others showed near zero correlation. The similarity in category selectivity in the latter cell pairs cannot be explained by similarity in stimulus selectivity. We have thus concluded that a local cell population contributes to object category discrimination by representing more than one independent features affiliated to the same categories. Then, how many features are represented by population of cells in a column? We decomposed their responses to 850 stimuli by independent component analysis, and found that there were 3-15 significant components with distribution of affinities significantly deviated from gaussian. These significant components were regarded as the features represented by the population of cells. Then, we compared the strength of each feature's affinity to individual stimuli between pairs of categories by using ROC analysis, and found that 2 to 7 of the features contributed to the categorical selectivity of population activities in each column.
O3-7-1-3
GIST記述子を利用したニューロン活動に基づく観察画像の復元
Image reconstruction from neural activities based on the GIST descriptor

○林隆介1,2, 倉重宏樹3
○Ryusuke Hayashi1,2, Hiroki Kurashige3
産総研・システム脳1, JST・さきがけ2, ATR・CNS3
System Neuro, AIST, Tsukuba, Japan1, PRESTO, JST, Tokyo, Japan2, CNS, ATR, Kyoto, Japan3

Introduction: Recent fMRI studies have shown that our visual experience can be reconstructed from evoked brain activities by modeling the function between the brain activities and visual input patterns (Nishimoto et al., 2011). The image reconstruction technique using fMRI, however, has a limitation to make brain decoding device small and to decode visual experience in real time. In the present study, we recorded spike activities of the neurons in the inferior temporal (IT) cortex using small electrodes and demonstrated that the GIST descriptor (a vector that describes orientation energies at different scales within an image) is useful to reconstruct the holistic patterns of the viewed images from the neural activities.
Method: We chronically implanted three multi-electrode arrays in the posterior, middle, and anterior part of the IT cortex of a macaque monkey and recorded multi-unit activities while the animal viewed one of 120 object images randomly presented.
Analysis:
The decoders for 120 object images were trained using ridge regression (L2 norm penalty) between the GIST descriptors (384 dimensions) of the images and the neural activities (190 electrodes) sampled every 100ms. We then estimated the GIST descriptors of the test data and chose top 16 images whose GIST descriptors were most similar as the estimated one from an image database (40000 images collected from Caltech 256 etc). Finally, the viewed images were reconstructed by averaging these top 16 images.
Results:
The results demonstrated that our simple decoder can reconstruct the holistic pattern of the viewed image. The quality of the reconstruction was best when the neural activities during 100-200ms after stimulus onset were used (The fitting of the GIST descriptors (correlation coefficient) was 0.31). To improve the quality of image reconstruction, however, it is critical to use more object-selective descriptors implemented in the recent computer vision studies using deep learning.
O3-7-1-4
サル下部側頭皮質36野とTE野間を伝達する記憶信号の電気生理学的解析
Across-areal analysis of mnemonic backward signal flow between area 36 and TE of macaques

○竹田真己1, 小谷野賢治1, 宮下保司1
○Masaki Takeda1, Kenji Koyano W1, Yasushi Miyashita1
東京大学大学院 医学系研究科 総合生理学1
Dept Physiol, Univ of Tokyo, Tokyo1

In this study, we conducted electrophysiological recordings simultaneously from area 36 (A36) and area TE when two macaque monkeys were performing the pair-association task. In each trial, after the presentation of the cue stimulus for 500 ms, two choice stimulus was presented with the delay of 1000 or 2000 ms. Monkeys were rewarded when they correctly chose the paired-associate of the cue stimulus. Extracellular recordings were conducted using tungsten electrodes and linear-array 16ch contact electrodes. Recorded neuronal signals were separated into single unit activities (SUA) and local field potentials (LFP). Current source density of LFP was calculated to estimate the channel located at the granular layer (Takeuchi et al., 2011). The location of recording sites was also identified on high resolution MR images in vivo using MRI detectable elgiloy deposit marking (Koyano et al., 2010) and was validated histologically. Spectral SUA-LFP coherence was calculated using the multi-taper method (Mitra and Pesaran, 1999). During the delay period, SUA in A36 was significantly coherent with LFP in TE especially in beta frequency band. This coherence was significantly larger than coherence of the opposite direction, SUA in TE and LFP in A36. In individual recordings, SUA in A36 was significantly coherent with LFP only at the infragranular layer, only at the supragranular layer or both at the infra- and supragranular layers of TE. In population, coherence was significantly larger both in the supra- and infragranular layers than in the granular layer. These results suggested that mnemonic signal flowed from A36 to TE in a layer specific manner.
上部に戻る 前に戻る