(A) simulation study on cortical long-range connections for cost-efficient visual function = 효율적인 시각 인지를 위한 시각 피질 장거리 연결의 기능에 대한 시뮬레이션 연구

Cited 0 time in webofscience Cited 0 time in scopus
  • Hit : 19
  • Download : 0
Both the human brain and recent deep neural networks (DNNs) successfully perform object recognition. However, the visual pathway consists of far fewer hierarchical stages compared to those in DNNs. The structural factor that achieves efficient visual function of the brain is still elusive. Here, we suggest that cortical long-range horizontal connections (LRCs) observed in the early visual cortex enable cost-efficient object recognition in shallow network. First, to validate our hypothesis, a shallow network with convergent feedforward connections and LRCs from neural connectivity data of tree shrew was designed. To investigate effect of LRCs for object recognition, the network was trained for CIFAR-10 image classification by changing network connectivity. We found that addition of LRCs to the shallow feedforward network significantly enhances the classification performance, even to comparable to much deeper network. Second, from gradient-based optimization by pruning the connections, we confirmed that LRCs could spontaneously emerge by balancing between minimizing total connection length and maximizing classification performance. After the optimization, a certain portion of long-range connections survived. Deletion of survived LRCs led to significant reduction of classification performance, which implies LRCs are important for object recognition. Third, to investigate how LRCs contribute to image processing, datasets with local or global structure of input images were generated by modifying handwritten digit dataset. We observed that the network with sparse LRCs and dense feedforward connections can consistently classify images with different local and global information. Lastly, we found that performance enhancement by LRCs is strongly correlated with small-worldness of network, and that this can explain the species-specific existence of LRCs in the visual cortex. In summary, we suggest that long-range horizontal connectivity may be a key factor allowing the visual cortex to implement cost-efficient object recognition under physical constraints.
Advisors
Paik, Se-Bumresearcher백세범researcher
Description
한국과학기술원 :바이오및뇌공학과,
Publisher
한국과학기술원
Issue Date
2020
Identifier
325007
Language
eng
Description

학위논문(석사) - 한국과학기술원 : 바이오및뇌공학과, 2020.2,[ii, 35 p. :]

Keywords

Long-range horizontal connection▼aNeural networks▼aObject recognition▼aCost-efficiency▼aSmall-world network; 수평적 장거리 연결▼a인공신경망▼a사물 인식▼a비용 효율성▼a작은 세상 신경망

URI
http://hdl.handle.net/10203/283836
Link
http://library.kaist.ac.kr/search/detail/view.do?bibCtrlNo=909917&flag=dissertation
Appears in Collection
BiS-Theses_Master(석사논문)
Files in This Item
There are no files associated with this item.

qr_code

  • mendeley

    citeulike


rss_1.0 rss_2.0 atom_1.0