Dynamic neuronal networks eciently achieve classication in robotic interactions with real-world objects Pakorn Uttayopas Xiaoxiao Cheng Udaya Bhaskar Rongala

2025-04-26 0 0 8.44MB 9 页 10玖币

侵权投诉

Dynamic neuronal networks eﬃciently achieve classiﬁcation in

robotic interactions with real-world objects

Pakorn Uttayopas, Xiaoxiao Cheng, Udaya Bhaskar Rongala,

Henrik J¨orntell, Etienne Burdet

Abstract

Biological cortical networks are potentially fully recurrent networks without any distinct output

layer, where recognition may instead rely on the distribution of activity across its neurons. Because

such biological networks can have rich dynamics, they are well-designed to cope with dynamical

interactions of the types that occur in nature, while traditional machine learning networks may

struggle to make sense of such data. Here we connected a simple model neuronal network (based

on the ’linear summation neuron model’ featuring biologically realistic dynamics (LSM), consisting

of 10 of excitatory and 10 inhibitory neurons, randomly connected) to a robot ﬁnger with multiple

types of force sensors when interacting with materials of diﬀerent levels of compliance. Scope: to

explore the performance of the network on classiﬁcation accuracy. Therefore, we compared the

performance of the network output with principal component analysis of statistical features of the

sensory data as well as its mechanical properties. Remarkably, even though the LSM was a very

small and untrained network, and merely designed to provide rich internal network dynamics while

the neuron model itself was highly simpliﬁed, we found that the LSM outperformed these other

statistical approaches in terms of accuracy.

1 Introduction

Here we aimed to use biologically relevant neuron models connected in a brain-like network structure to

study its potential to achieve input separation in a robotic system interacting with real-world objects.

The model network was inspired by local cortical networks in its recursive structure in principle, though

with much fewer neurons and without the ambition to precisely mimick any assumed speciﬁc network

structure. The aim was to explore if the inherent dynamic properties in such networks in themselves

were enough to achieve eﬃcient object classiﬁcation.

Our model system is reminiscent of Reservoir Computing networks (i.e. Gauthier et al 2020 Nature

Communications), but our neurons have state memory, i.e. dynamics, which are biologically relevant.

Moreover, the population of neurons are split into excitatory and inhibitory neurons. Combined with

the neuronal output thresholding, i.e. imparting nonlinearity to the networks when inhibition drives

the neurons below their thresholds, and combined with biologically relevant conduction delays, this

setting creates extraordinarily rich network dynamics.

Motivation for: what would be required in the robotics design to explore the questions we set out

to explore? How well could we live up to those requirements with the robotics system used?

2 Methods

2.1 Neuron model

The neuron model used in this work was a non-spiking Linear Summation Model (LSM) with an

additional dynamic leak component [1]. LSM aims to capture the important characteristics of a

Hodgkin-Huxley (H-H) conductance-based model [2]. The LSM describes the activity dynamics {aj(t)}

of “cortical neurons” arising from the weighted activity of other cortical neurons (with inhibitory αi<0

and excitatory αi>0 projection) as well as sensory neurons {bk}:

τdaj

dt =−a+

j(t) + Pi6=jαiai(t) + Pkβkbk(t)

Pi6=j|αi|ai(t) + Pk|βk|bk(t) + γ, a+

j=(ajaj>0

0aj≤0(1)

arXiv:2210.06303v2 [q-bio.NC] 11 Nov 2022

Importantly, the forgetting of each cortical neuron activity −a+

j(t) ensures convergence, at a speed

regulated by the time constant τ. The denominator weights the inﬂuence of other neurons, with a

factor γ > 0 avoiding divergence when they are not active.

2.2 Neuron models conﬁguration

2.2.1 Network Connectivity

In this work, the network connectivity is comprised of 10 excitatotry nodes (ENs, blue markers) and

10 inhibitory nodes (INs, red markers), totalling 20 neurons. All nodes randomly received the external

inputs which are force, torque and vibration from object-robot interaction. ENs are bi-directionally

connected with all other nodes, but the central one. However, INs are connected only with ENs, but

not within INs (see Fig 1.).

4 connectivity conﬁgurations {full connectivity, 75% connectivity, 50% connectivity and 25% con-

nectivity}were considered. In full connectivity, all nodes within the network are fully connected as

mentioned previously (see Fig 1.a). In other conﬁgurations, we randomly removed 25%, 50% and 75%

connections from the full connectivity (see Fig 1.b-d respectively). Note that full connectivity was

used for the rest of network simulations, otherwise speciﬁed.

Figure 1: Visualization of the network connections at the diﬀerent connectivity levels. In (a)-(d), red

nodes are inhibitory neurons and blue nodes are excitatory neurons. Full connectivity (a) implied that

every single node (neuron) was connected to all other nodes, but the weight of each connection was

varied according to random patterns. Reduced connectivity was achieved by removing connections

randomly to (b) 75% (c) 50% (c) 25% connectivity.

2.2.2 Static and Dynamic leak

The LSM neurons have both a static leak and a dynamic leak constant [1]. The static leak constant,

kwas ﬁxed at 2 as it only acts as a normalization factor of the neuron activity, without aﬀecting the

network dynamics. On the other hand, we consider the dynamic leak time constant at τDyn = 1/100

as this constant acts as a low-pass ﬁltering factor for neuron activity and thereby impact the network

dynamics.

2.2.3 Conduction Delays

An axonal delays between two neurons were also applied in the network. The conduction delays in

communication among neurons were uniformly distributed randomised in range of [0,1] ms. These

values of conduction delays in the simulations denotes as timesteps of signal lagged between two

neurons in the network.

2.2.4 Input Weights

All external inputs from ENs are weighted. The weights were generated as normal distributions with

a distribution mean (µ) of 0.3 and standard deviation (σ) of 0.5.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

Dynamicneuronalnetworksecientlyachieveclassicationinroboticinteractionswithreal-worldobjectsPakornUttayopas,XiaoxiaoCheng,UdayaBhaskarRongala,HenrikJorntell,EtienneBurdetAbstractBiologicalcorticalnetworksarepotentiallyfullyrecurrentnetworkswithoutanydistinctoutputlayer,whererecognitionmayinsteadr...

展开>> 收起<<

Dynamic neuronal networks eciently achieve classication in robotic interactions with real-world objects Pakorn Uttayopas Xiaoxiao Cheng Udaya Bhaskar Rongala.pdf

共9页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Dynamic neuronal networks eciently achieve classication in robotic interactions with real-world objects Pakorn Uttayopas Xiaoxiao Cheng Udaya Bhaskar Rongala

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: