Sparse long-range connections in visual cortex for cost-efficient small-world networks

Abstract

9The brain performs visual object recognition using much shallower hierarchical stages than artificial deep 10 neural networks employ. However, the mechanism underlying this cost-efficient function is elusive. Here, 11 we show that cortical long-range connectivity(LRC) may enable this parsimonious organization of circuits 12 for balancing cost and performance. Using model network simulations based on data in tree shrews, we 13 found that sparse LRCs, when added to local connections, organize a small-world network that 14 dramatically enhances object recognition of shallow feedforward networks. We found that optimization of 15 the ratio between LRCs and local connections maximizes the small-worldness and task performance of 16 the network, by minimizing the total length of wiring needed for integration of the global information. We 17 also found that the effect of LRCs varies by network size, which explains the existence of species-specific 18 LRCs in mammalian visual cortex of various sizes. Our results demonstrate a biological strategy to 19 achieve cost-efficient brain circuits. 20 21 Highlights 22 • Long-range connections (LRCs) enhance the object recognition of shallow networks 23 • Sparse LRCs added to dense local connections organize a small-world type network 24 • Small-worldness of networks modulates the balance between performance and wiring cost 25 • Distinct LRCs in various species are due to the size-dependent effect of LRCs 26 Significance statement 27The hierarchical depth of the visual pathway in the brain is constrained by biological factors, whereas 28 artificial deep neural networks consist of super-deep structures (i.e., as deep as computational power 29 allows). Here, we show that long-range horizontal connections (LRCs) observed in mammalian visual 30 cortex may enable shallow biological networks to perform cognitive tasks that require deeper artificial 31 structures, by implementing cost-efficient organization of circuitry. Using model simulations based on 32 anatomical data, we found that sparse LRCs, when added to dense local circuits, organize "small-world" 33 type networks and that this dramatically enhances image classification performance by integrating both 34 local and global components of visual stimulus. Our findings show a biological strategy of brain circuitry 35 to balance sensory performance and wiring cost in the networks. 36 37 One sentence summary 38 Cortical long-range connections organize a small-world type network to achieve cost-efficient functional 39 circuits under biological constraints 40 41Recently, computer vision techniques for object recognition using deep neural networks (DNNs) have 43 been remarkably developed. These now show performance comparable to that of humans 1-5 , implying 44 that deep artificial neural network models can provide collective knowledge of visual processing in the 45 brain 6-8 . However, there exist fundamental differences between the two kinds of systems, particularly 46 regarding the structure required for visual feature abstraction. ...

Invariance of object detection in untrained deep neural networks

“…Currently, DNN models, which have a biologically inspired hierarchical structure, provide an effective approach for investigating functions in the brain ( Paik and Ringach, 2011 ; DiCarlo et al, 2012 ; Yamins and DiCarlo, 2016 ; Sailamul et al, 2017 ; Baek et al, 2020 , 2021 ; Jang et al, 2020 ; Kim et al, 2020 , 2021 ; Park et al, 2021 ; Song et al, 2021 ). Several studies have reported that DNNs trained to natural images can predict the neural responses of the monkey inferior temporal cortex (IT) ( Cadieu et al, 2014 ; Yamins et al, 2014 ), known as the area for object recognition.…”

“…A model study using a biologically inspired deep neural network (DNN) ( Krizhevsky et al, 2012 ; Simonyan and Zisserman, 2015 ) has been suggested as an effective approach to this problem ( Paik and Ringach, 2011 ; DiCarlo et al, 2012 ; Yamins and DiCarlo, 2016 ; Sailamul et al, 2017 ; Baek et al, 2020 , 2021 ; Jang et al, 2020 ; Kim et al, 2020 , 2021 ; Park et al, 2021 ; Song et al, 2021 ). DNNs, which consist of a stack of feedforward projections inspired by the hierarchical structure of the visual pathway, can be used as simplified model to investigate various visual functions.…”

Invariance of object detection in untrained deep neural networks

“…Currently, deep neural network (DNN) models, which have a biologically inspired hierarchical structure, provide an effective approach for investigating functions in the brain (Paik and Ringach, 2011;DiCarlo et al, 2012;Yamins and DiCarlo, 2016;Sailamul et al, 2017;Baek et al, 2020Baek et al, , 2021Jang et al, 2020;Kim et al, 2020Kim et al, , 2021Park et al, 2021;Song et al, 2021). Several studies have reported that DNNs trained to natural images can predict the neural responses of the monkey inferior temporal cortex (IT) (Cadieu et al, 2014;Yamins et al, 2014), known as the area for object recognition.…”

“…A model study using a biologically inspired deep neural network (DNN) (Krizhevsky et al, 2012; Simonyan and Zisserman, 2015) has been suggested as an effective approach to this problem (Paik and Ringach, 2011; DiCarlo et al, 2012; Yamins and DiCarlo, 2016; Sailamul et al, 2017; Baek et al, 2020, 2021; Jang et al, 2020; Kim et al, 2020, 2021; Park et al, 2021; Song et al, 2021). DNNs consist of a stack of feedforward projections inspired by the hierarchical structure of the visual pathway and can be used as simplified model to investigate various visual functions.…”