Utilizing Explainable AI for improving the Performance of Neural Networks Huawei Sun13 Lorenzo Servadei13 Hao Feng13 Michael Stephan12 Robert Wille3 Avik Santra1

2025-05-06 0 0 631.17KB 8 页 10玖币

侵权投诉

Utilizing Explainable AI for improving the

Performance of Neural Networks

Huawei Sun1,3, Lorenzo Servadei1,3, Hao Feng1,3, Michael Stephan1,2, Robert Wille3, Avik Santra1

1Inﬁneon Technologies AG, Neubiberg, Germany

2Friedrich-Alexander-University Erlangen-Nuremberg, Erlangen, Germany

3Technical University of Munich, Munich, Germany

E-mail: {huawei.sun, lorenzo.servadei, avik.santra}@inﬁneon.com

{hao.feng, robert.wille}@tum.de

{michael.stephan}@fau.de

Abstract—Nowadays, deep neural networks are widely used in

a variety of ﬁelds that have a direct impact on society. Although

those models typically show outstanding performance, they have

been used for a long time as black boxes. To address this,

Explainable Artiﬁcial Intelligence (XAI) has been developing as

a ﬁeld that aims to improve the transparency of the model and

increase their trustworthiness. We propose a retraining pipeline

that consistently improves the model predictions starting from

XAI and utilizing state-of-the-art techniques. To do that, we use

the XAI results, namely SHapley Additive exPlanations (SHAP)

values, to give speciﬁc training weights to the data samples. This

leads to an improved training of the model and, consequently,

better performance. In order to benchmark our method, we

evaluate it on both real-life and public datasets. First, we

perform the method on a radar-based people counting scenario.

Afterward, we test it on the CIFAR-10, a public Computer Vision

dataset. Experiments using the SHAP-based retraining approach

achieve a 4% more accuracy w.r.t. the standard equal weight

retraining for people counting tasks. Moreover, on the CIFAR-10,

our SHAP-based weighting strategy ends up with a 3% accuracy

rate than the training procedure with equal weighted samples.

Index Terms—Radar Sensors, Explainable AI, Deep Learning,

SHapley additive exPlanations

I. INTRODUCTION

Various application areas have been positively affected by

the recent advances of Artiﬁcial Intelligence (AI) and Machine

Learning (ML). Among them, ﬁelds such as autonomous

driving [1], [2], health tech [3], [4] and robotics [5], [6]

heavily rely on the processing of ML algorithms onto a set

of different sensors. These approaches are typically based

on computationally intensive Deep Learning (DL) strategies,

which involve training millions, or even billions of parameters

to perform a speciﬁc task. Although the out-coming results

show high performance, a major problem occurs: As a neural

network gets deeper and deeper, it is also becoming more

complex and thus challenging to be interpreted. To this end,

a neural network is often considered a black box: Even if the

model correctly predicts the given speciﬁc input, it is difﬁcult

to explain what causes the correct prediction. This property,

in turn, reduces the trustworthiness of the outcome.

In order to improve a DL system, it is necessary to under-

stand its weaknesses and shortcomings [7]. Approaching this,

XAI focuses on improving the transparency of ML technolo-

gies and increasing their trust. When a model’s predictions

are incorrect, explanatory algorithms can aid in tracing the

underlying reasons and phenomenon. XAI has been researched

for several years, and lots of work has been done in ﬁelds

such as Computer Vision (CV) [8], [9], [10] and Natural

Language Processing (NLP) [11], [12]. These algorithms

mainly generate attention maps, which help to highlight the

critical area/words in classifying images or during language

translation. Nevertheless, nowadays DL is widely applied in

less conventional application ﬁelds: For example, radar-based

solutions for tasks such as counting people [13], identifying

gestures [14], and tracking [15], as shown in this contribution

[16]. Although the advancements mentioned above success-

fully solve radar-based problems, explaining DL models for

radar signals is still a challenging topic. Additionally, most

XAI algorithms analyze the predictions from a well-trained

model, thus focusing only on the explanatory part. To this end,

a few research contributions move forward by utilizing the

results from XAI for secondary tasks. Layer-wise Relevance

Propagation (LRP), for example, is used for adaptive learning

rate during training in [17], and in [18] the authors prune Deep

Neural Networks (DNNs) and quantize the weights mainly by

Deep Learning Important FeaTures (DeepLIFT). However, to

the best of our knowledge, XAI has not yet been used to

process the dataset and improve the network performance. In

this paper, we ﬁrst adapt our method to a real-life use case:

Radar-based people counting. Afterward, we show promising

results on the CIFAR-10 dataset [19] to further underlying the

approach’s generality.

Radar-based people counting is a signiﬁcant application

with high privacy preservation and weather condition indepen-

dence compared to camera-based people counting. However,

the algorithms often underperform the state-of-the-art com-

puter vision methods, and radar data often has the limitation

of low-resolution and room dependency [20], [21]. To this end,

many solutions have been implemented which use DL for this

task in different scenarios [13], [22]. Although performant,

those solutions do not consider how the network obtains the

prediction and which features are essential to explain the

outcome of the task.

This paper introduces a retraining pipeline, which adopts

arXiv:2210.04686v1 [cs.LG] 7 Oct 2022

the SHAP values for improving the model performance. The

proposed approach shows convincing outcomes in both public

and real-life datasets. In the radar-based people counting

scenario, on the one hand, we explain the neural network’s

prediction of radar signals. This is the ﬁrst time in the

literature that XAI algorithms are applied to radar-based input

neural networks. On the other hand, we propose a retraining

pipeline that adds sample weights generated from the SHAP

[23] results. This further improves the performance of the

people counting network. To show this, we execute several

experiments highlighting our method’s beneﬁt. We ﬁrst apply

our weighting strategy to our radar-based people counting task.

Compared with the equal weighting method, our SHAP-based

method ends up with an increase of 4%. When we apply the

method to CIFAR-10, our approach outperforms default equal

weighting retraining with a 3% more accuracy.

II. BACKGROUND AND MOTIVATION

This section ﬁrst describes the status of explainable AI

and its application ﬁelds. Afterward, we introduce the funda-

mentals of mmWave Frequency-Modulated Continuous-Wave

(FMCW) radar.

A. Explainable Artiﬁcial Intelligence

Several class activation mapping (CAM) based XAI algo-

rithms such as CAM [8] and gradient class activation mapping

(Grad-CAM) [9] are used, in the literature, to explain well-

trained CNN architecture in image classiﬁcation tasks. They

typically work by generating saliency maps through a linear

combination of activation maps that highlight the model’s

attention area. Although those methods are widely-spread,

there is still a lack of research on using XAI algorithms to

explain models with radar signals as input. In fact, on the one

hand, features in radar signals are more difﬁcult to understand

for humans than in images. On the other hand, unlike image

RGB channels, radar signals are typically represented by

distinct information, such as Macro- and Micro-Doppler maps

which cannot be stacked and treated as one single image. This

contrasts with the usual practice in CV, where methods such as

CAM-based approaches only generate a single saliency map

for features of each input sample. Therefore, a major question

is: how to utilize an efﬁcient XAI method adaptable to the

heterogeneous information contained in the input. Recently,

SHAP[23] has been used in domains such as text classiﬁcation

[24] or analysis of time-series data [25], as an additive feature

attribution approach, which can explain multiple information

in the input at the same time. This exactly adapts to the

problem at hand.

Additive feature attribution method Additive feature attri-

bution method follows:

g(z0) = φ0+

i=1

φiz0

i(1)

where z0∈ {0,1}is a binary vector whose entries are the exis-

tence of the corresponding input feature and Mis the number

of simpliﬁed input features. φiindicates the importance of

the ith feature and φ0is the baseline explanation. In order

to explain a complex model f, such as a neural network, we

can use a more straightforward explanation model ginstead.

For a given input x,f(x)denotes the prediction output. A

simpliﬁed input x0can be restored to the input xthrough a

mapping function x=hx(x0). When the binary vector z0≈x0,

additive feature attribution methods ensure g(z0)≈fc(hx(x0))

where cis the target class of given input x.

SHAP values SHAP values are the feature attributions of the

explanation model, which obeys Eq. 1 and are formulated as

follows:

φi=X

z0⊂x0

(M− |z0|)!(z0−1)!

M![fc(hx(z0)) −fc(hx(z0\i))]

(2)

where |z0|denotes the number of non-zero entries in z0and

z0⊂x0represents all z0vectors where the non-zero entries

are a subset of the non-zero entries in x0. Additionally, z0\i

denotes setting zito zero. In this way, taking image-based

input as an example, for every single model input, it can

generate one SHAP map in the same shape as the input sample,

where the pixel value denotes attribution of the corresponding

pixel from the input.

B. Introduction of mmWave FMCW Radar Sensor

FMCW radar sensors typically operate by transmitting a

sequence of modulated frequency chirp signals with a short

ramp time and delays between them. A chirp sequence with

a ﬁxed number of chirps is usually deﬁned as one frame

and repeated with the frame repetition time [13]. The chirp

signals reﬂected by the targets are received, mixed with the

transmitted signal, and ﬁltered, to generate the Intermedi-

ate Frequency (IF) signal, which is then digitized by the

Analog to Digital Converter (ADC) for task-dependent pre-

processing. For example, range and Doppler information can

be acquired by taking the Fast Fourier Transform (FFT) along

the respective axes of the data in one frame. Using multiple

receiving or transmitting antennas also allows the estimation

of the Direction Of Arrival (DOA) from the DOA-dependent

time delay of the received signal across the antennas. This

makes FMCW radar the most commonly used since it avoids

complicated pre-processing and saves energy at the same time.

Therefore, it is an ideal sensor for many simple tasks, such

as people counting and activity classiﬁcation, without losing

users’ data privacy.

III. APPROACH

In this section, we ﬁrst describe the radar data preprocessing

method. Then, we illustrate the radar data augmentation meth-

ods. We also introduce a stabilized architecture for learning

robust embedding vectors and label predictions. Ultimately,

we focus on our retraining procedure with different weighting

methods calculated from SHAP values and probability vector

predictions.

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：

UtilizingExplainableAIforimprovingthePerformanceofNeuralNetworksHuaweiSun1;3,LorenzoServadei1;3,HaoFeng1;3,MichaelStephan1;2,RobertWille3,AvikSantra11InneonTechnologiesAG,Neubiberg,Germany2Friedrich-Alexander-UniversityErlangen-Nuremberg,Erlangen,Germany3TechnicalUniversityofMunich,Munich,GermanyE-...

展开>> 收起<<

Utilizing Explainable AI for improving the Performance of Neural Networks Huawei Sun13 Lorenzo Servadei13 Hao Feng13 Michael Stephan12 Robert Wille3 Avik Santra1.pdf

共8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！

Utilizing Explainable AI for improving the Performance of Neural Networks Huawei Sun13 Lorenzo Servadei13 Hao Feng13 Michael Stephan12 Robert Wille3 Avik Santra1

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: