Neural Co-Processors for Restoring Brain Function Results from a Cortical Model of Grasping Matthew J Bryan1 Linxing Preston Jiang123 Rajesh P N

2025-05-02 0 0 3.67MB 45 页 10玖币

侵权投诉

Neural Co-Processors for Restoring Brain Function:

Results from a Cortical Model of Grasping

Matthew J Bryan1, Linxing Preston Jiang1,2,3, Rajesh P N

Rao1,2,3

1Neural Systems Laboratory, Paul G. Allen School of Computer Science &

Engineering, University of Washington, Seattle, WA, USA

2Center for Neurotechnology, University of Washington, Seattle, WA, USA

3Computational Neuroscience Center, University of Washington, Seattle, WA, USA

E-mail: {mmattb,prestonj,rao}@cs.washington.edu

March 2023

Abstract.

Objective A major challenge in designing closed-loop brain-computer

interfaces (BCIs) is ﬁnding optimal stimulation patterns as a function of ongoing neural

activity for diﬀerent subjects and diﬀerent objectives. Traditional approaches, such as

those currently used for deep brain stimulation (DBS), have largely followed a manual

trial-and-error strategy to search for eﬀective open-loop stimulation parameters, a

strategy that is ineﬃcient and does not generalize to closed-loop activity-dependent

stimulation. Approach To achieve goal-directed closed-loop neurostimulation, we

propose the use of brain co-processors, devices which exploit artiﬁcial intelligence (AI)

to shape neural activity and bridge injured neural circuits for targeted repair and

restoration of function. Here we investigate a speciﬁc type of co-processor called a

“neural co-processor” which uses artiﬁcial neural networks (ANNs) and deep learning to

learn optimal closed-loop stimulation policies. The co-processor adapts the stimulation

policy as the biological circuit itself adapts to the stimulation, achieving a form of

brain-device co-adaptation. Here we use simulations to lay the groundwork for future

in vivo tests of neural co-processors. We leverage a previously published cortical

model of grasping, to which we applied various forms of simulated lesions. We used

our simulations to develop the critical learning algorithms and study adaptations to

non-stationarity in preparation for future in vivo tests. Main results Our simulations

show the ability of a neural co-processor to learn a stimulation policy using a supervised

learning approach, and to adapt that policy as the underlying brain and sensors change.

Our co-processor successfully co-adapted with the simulated brain to accomplish the

reach-and-grasp task after a variety of lesions were applied, achieving recovery towards

healthy function in the range 75-90%. Signiﬁcance Our results provide the ﬁrst proof-

of-concept demonstration, using computer simulations, of a neural co-processor for

adaptive activity-dependent closed-loop neurostimulation for optimizing a rehabilitation

goal after injury. While a signiﬁcant gap remains between simulations and in vivo

applications, our results provide insights on how such co-processors may eventually

be developed for learning complex adaptive stimulation policies for a variety of neural

rehabilitation and neuroprosthetic applications.

arXiv:2210.11478v2 [q-bio.NC] 20 Mar 2023

Neural Co-Processors 2

Keywords: brain-computer interface, brain-machine interface, neurostimulation,

neuromodulation, neural co-processor, AI, machine learning, deep learning, neural

networks, computational models

Neural Co-Processors 3

1. Introduction

Brain-computer interfaces (BCIs) have made signiﬁcant advances over the last several

decades, leading to the control of a wide variety of virtual and physical prostheses

through neural signal decoding [

]. Separately, advances in stimulation techniques

and modeling have allowed us to probe neural circuit dynamics (e.g. [

]) and learn to

better drive neural circuits towards desired target dynamics by encoding and delivering

information through stimulation [

]. Bi-directional BCIs (BBCIs)

allow stimulation to be conditioned on decoded brain activity and encoded sensor data

for applications such as real-time, ﬁne-grained control of neural circuits and prosthetic

devices (e.g., [14]).

Motivated by these advances, we investigate here a ﬂexible framework for combining

encoding and decoding using “neural co-processors” [

], a type of brain co-processor

[

]. Neural co-processors leverage artiﬁcial neural networks (ANNs) and deep learning

to compute optimal closed-loop stimulation patterns. The approach can be used to

not only drive neural activity toward desired activity regimes, but also to achieve task

goals external to the subject, such as ﬁnding closed-loop stimulation patterns for motor

cortical neurons for restoring the ability to reach and grasp an object. Likewise, the

framework generalizes to stimulation based on both brain activity and external sensor

measurements, e.g., from cameras or light detection and ranging (LIDAR) sensors, in

order to restore perception (e.g., cortical visual prosthesis) or incorporate feedback for

real-time prosthetic control (see [16] for details).

The co-processor framework also allows co-adaptation with biological circuits in

the brain by updating its stimulation policy, while the brain updates its own response

to the stimulation via adaptation and neural plasticity, or modiﬁes its response due

to other reasons. The co-processor could potentially optimize its outputs for a desired

optimization function continually in the presence of signiﬁcant non-stationarities in the

brain.

Here, to lay the groundwork for future in vivo tests of the co-processor framework,

we use computer simulations to explore how co-processors can be trained to restore

lost function and how they can adapt to non-stationarities. We demonstrate a neural

co-processor that restores movement in a computational model of cortical networks

involved in controlling a limb, after a simulated stroke aﬀects the ability to use that

limb. Our demonstration combines both components of a neural co-processor [15]:

•

An emulation model based on ANNs, which learns a mapping from stimulation

and current neural activity to output variables such as task performance (or future

neural activity).

•

An artiﬁcial intelligence (AI) “agent” based on ANNs which learns the best closed-

loop activity-dependent stimulation to apply in real time to optimize a given task.

Neural Co-Processors 4

2. Background

Signiﬁcant advances have been made in understanding and modeling the eﬀects of

electrical and other forms of neural stimulation on the brain. Researchers have

explored how information can be biomimetically or artiﬁcially encoded and delivered via

stimulation to neuronal networks in the brain and other regions of the nervous system

for auditory [

], visual [

], proprioceptive [

], and tactile [

] perception.

Advances have also been made in modeling the eﬀects of stimulation over large scale,

multi-region networks, and across time [

]. Some models can additionally adapt to

ongoing changes in the brain, including changes due to the stimulation itself [

]. For our

simulations described below, we use a stimulation model, not unlike those cited above,

which seeks to account for both network dynamics and non-stationarity. In addition

to training the model to have a strong ability to predict the eﬀect of stimulation, we

additionally adapt it to be useful for learning an optimal stimulation policy, a property

distinct from predictive power alone.

Researchers have also explored both open- and closed-loop stimulation protocols

for treating a variety of disorders. Open loop stimulation has been eﬀective in treating

Parkinson’s Disease (PD) [

], as well as various psychiatric disorders [

]. In

research more directly related to our work, Khanna et al. [

] investigated the use of

open loop stimulation in restoring dexterity after a lesion in nonhuman primate’s (NHP)

motor cortex. The authors demonstrate that the use of low-frequency alternating current,

applied epidurally, can improve grasp performance.

While open loop stimulation techniques have yielded clinically useful results, results

in many domains have been mixed, such as in visual prostheses [

], and in invoking

somatosensory feedback [

]. We believe this is due to the stimulation not being

conditioned on the ongoing dynamics of the neural circuit being stimulated. From

moment to moment and throughout the day, a neuronal circuit in the brain can be

expected to respond diﬀerently even when the same stimulation parameters are used,

due to the multitude of diﬀerent external and internal inputs inﬂuencing the circuit’s

ongoing activity. Stimulation therefore needs to be closed-loop, i.e. proactively adapted

in response. This need is even greater over longer time scales as the eﬀects of plasticity,

changes in clinical conditions, and ageing change the dynamics and connectivity of

the brain. Closed-loop stimulation may also provide means to better regulate the

energy use of an implanted stimulator, allowing it to intelligently regulate when to

apply stimulation, in order to preserve implant battery life [

]. Another beneﬁt is that

closed-loop stimulation oﬀers an opportunity to minimize the side-eﬀects of stimulation,

through real time regulation of the stimulation parameters, such as in the use of deep

brain stimulation (DBS) in PD patients [

]. In recent years, closed-loop stimulation

has been used to aid in learning new memories after some impairment [

], to replay

visually-invoked activations [

], and for optogenetic control of a thalamocortical circuit

[29], among others.

A major open question is: how does one leverage closed-loop stimulation for real-time

Neural Co-Processors 5

co-adaptation with the brain to accomplish an external task such as restoration of a lost

function? “Co-adaptation” here refers to the ability of a BCI to adapt its stimulation

regime to ongoing changes in neural circuits in the brain, and to adapt with the brain to

accomplish the external task (e.g., grasping). The neural co-processor we present here

provides one potential approach to accomplishing this goal. Through the use of deep

learning, a neural co-processor co-adapts its AI, which controls stimulation, in synch

with the biological circuits in the brain.

For a neurologically complex task such as grasping, it is unlikely that there exists a

ﬁxed real-time controller which can be identiﬁed a priori for stimulating the (potentially

impaired) neural circuits involved in the task. This is due in large part to the variability

in the placement and performance of sensors and stimulators in diﬀerent brains, as well

as variability in brain structure and function between subjects. The most plausible path

to implementing a real-time controller is therefore to allow the device to adapt to the

subject, and to the long-term changes in their brain activities, and variability in the

sensors, stimulators and hardware. Our proposed neural co-processors seek to accomplish

such adaptation through ANNs and deep learning, and a particular training paradigm

described below.

2.1. Simulation as a Way to Gain Insights Prior to in vivo Experiments

To gain insights into neural co-processors before testing them in vivo, we investigated a

number of crucial design elements through the use of a previously published model by

Michaels et al. [

] of the cortical areas involved in grasping; the model, based on multiple

recurrent neural networks, is inspired by cortical anatomy and was ﬁt to data from

nonhuman primates performing grasping tasks. Using this cortical model as a “simulated

brain” allowed us to rapidly iterate through diﬀerent design and training methods to

demonstrate key properties of the neural co-processor framework. These insights will

help guide the co-processor training methodologies and experimental design for future in

vivo experiments. Additionally, a commonly-accepted maxim of animal experimentation

is that animal use should be narrowly tailored to answering questions which cannot be

answered in other ways. Consider, for example, “the 3Rs alternatives” approach to the

use of animal experiments [

]. In our case, we leverage computer simulations for initial

investigations into co-processor design, gathering evidence in preparation for future in

vivo experiments. In the Discussion section, we explore potential paths for translation of

this work to in vivo experiments.

A previous example of such a simulation approach is the work of Dura-Bernal et

al. [

]. In this work, the authors used a simulated spiking neural network to train

a stimulation agent. Their stimulation agent sought to restore the network’s control

of a simulated arm to reach a target, after a simulated lesion was applied. Similar

to our approach, the authors simulated lesions by eﬀectively removing parts of their

simulated network, or by removing connections between parts of the network. As the

authors point out, there exists only limited ability to probe a neural circuit in vivo in

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：

NeuralCo-ProcessorsforRestoringBrainFunction:ResultsfromaCorticalModelofGraspingMatthewJBryan1,LinxingPrestonJiang1;2;3,RajeshPNRao1;2;31NeuralSystemsLaboratory,PaulG.AllenSchoolofComputerScience&Engineering,UniversityofWashington,Seattle,WA,USA2CenterforNeurotechnology,UniversityofWashington,Seattl...

展开>> 收起<<

Neural Co-Processors for Restoring Brain Function Results from a Cortical Model of Grasping Matthew J Bryan1 Linxing Preston Jiang123 Rajesh P N.pdf

共45页,预览5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！

Neural Co-Processors for Restoring Brain Function Results from a Cortical Model of Grasping Matthew J Bryan1 Linxing Preston Jiang123 Rajesh P N

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: