Human-Centric AI for Trustworthy IoT Systems With Explainable Multilayer Perceptrons
Human-Centric AI for Trustworthy IoT Systems With Explainable Multilayer Perceptrons
复制标题
DOI：
10.1109/access.2019.2937521
复制
发表时间：
2019-01-01
期刊：
影响因子：
3.9
通讯作者：
Lloret, Jaime
中科院分区：
文献类型：
--
作者：
Garcia-Magarino, Ivan;Muttukrishnan, Rajarajan;Lloret, Jaime
物联网（IoT）广泛使用人工智能（AI）技术进行数据分析，以便从用户操作中学习，支持决策，跟踪用户的相关方面，并在适当时通知某些事件。然而，大多数人工智能技术都是基于数学模型的，一般公众很难理解，因此大多数人将基于人工智能的技术作为一个黑匣子，他们最终会根据个人经验开始信任它。本文提出了在物联网中使用人工智能的一个步骤，并在以人为中心的人工智能领域提出了一种新的方法，用于生成关于神经网络（特别是多层感知器）从物联网环境中学习的知识的解释。更具体地说，这项工作提出了两种技术的基础上分析人工神经元的权重，另一种技术旨在解释每个估计的基础上分析的训练情况。这种方法已经在智能物联网厨房的背景下进行了说明，该厨房基于每餐使用的食物检测用户抑郁症，并为此使用模拟器。结果显示，大多数自动生成的解释在这种情况下是有意义的（即97.0%），并且执行时间很低（即1.5 ms或更低），即使考虑到常见的配置，每个隐藏层的神经元数量（最多20个），隐藏层的数量（最多20个）和训练案例的数量（最多4，000个）也是独立变化的。
Internet of Things (IoT) widely use analysis of data with artificial intelligence (AI) techniques in order to learn from user actions, support decisions, track relevant aspects of the user, and notify certain events when appropriate. However, most AI techniques are based on mathematical models that are difficult to understand by the general public, so most people use AI-based technology as a black box that they eventually start to trust based on their personal experience. This article proposes to go a step forward in the use of AI in IoT, and proposes a novel approach within the Human-centric AI field for generating explanations about the knowledge learned by a neural network (in particular a multilayer perceptron) from IoT environments. More concretely, this work proposes two techniques based on the analysis of artificial neuron weights, and another technique aimed at explaining each estimation based on the analysis of training cases. This approach has been illustrated in the context of a smart IoT kitchen that detects the user depression based on the food used for each meal, using a simulator for this purpose. The results revealed that most auto-generated explanations made sense in this context (i.e. 97.0%), and the execution times were low (i.e. 1.5 ms or lower) even considering the common configurations varying independently the number of neurons per hidden layer (up to 20), the number of hidden layers (up to 20) and the number of training cases (up to 4,000).