索尼神经网络控制台Windows版本进行了更新，主要更新包括自定义单元功能、导入预训练模型作为单元功能以及添加了与XAI相关的插件。自定义单元功能允许用户将自定义的网络结构注册为组件并添加到网络中，方便复用常用的网络架构。导入预训练模型功能支持导入nnp、onnx和pb文件作为单元，简化迁移学习过程。此外，还添加了四个与可解释AI相关的插件，例如SGD Influence、Face Evaluation、LIME (batch)和LIME (tabular)，用于评估输入图像对识别结果的影响、测量人脸肤色、可视化影响分类结果的图像区域以及解释分类结果等。

🤔 **自定义单元功能:** 新增功能允许用户将自定义定义的单元注册为组件，并通过简单操作将其添加到网络中。用户可以将常用的网络结构保存为项目文件，然后将其注册为自定义单元，方便后续重复使用。

🔄 **导入预训练模型作为单元:** 新增功能允许用户将预训练的模型（如nnp、onnx和pb文件）导入为单元。用户可以通过右键菜单选择导入预训练模型文件，将其导入到当前项目中，方便进行迁移学习。

💡 **添加XAI相关插件:** 新增了四个与可解释AI相关的插件，包括SGD Influence、Face Evaluation、LIME (batch)和LIME (tabular)。这些插件分别用于评估输入图像对识别结果的影响、测量人脸肤色、可视化影响分类结果的图像区域以及解释分类结果等，有助于理解模型的决策过程。

We have updated the Neural Network Console Windows today.
This post will introduce the major updates.

・ Custom unit function
・ Import as unit function of pre-trained model
・ Addition of XAI related plug-ins

 

1. Custom unit function

A function has been added that allows you to register your own defined unit as a component and add it to the network with a simple operation.

The usage of the custom unit function is as follows.

1. Edit the network structure you want to reuse as a unit and save it as a project file.

2. From the right-click menu “Register Custom Unit” in the component list, select the saved project file.

The selected project will be registered in the Custom Unit category at the bottom of the component list.

3. Add the registered custom unit to the network in the same way as adding a normal layer.

You can add the project saved in step 1 to the current network as a unit.

The custom unit function makes it easier to reuse your frequently used network architecture.

 

2. Import as Unit function of pre-trained model

A function to import pre-trained models such as nnp, onnx, and pb files as a unit has been added.

From the right-click menu on the Edit tab, select Import, “Import, nntxt, nnp, ONNX as Unit” and select the pre-trained model file.
The selected pre-trained model will be imported as a unit into the current project.

This feature makes it easier to perform transfer learning using a pre-trained model.

 

3. Addition of XAI related plugins

The following four plug-ins have been added as Explainable AI (XAI) related plug-ins.

SGD Influence
Using a method called SGD Influence [1], the influence of the input images on recognition result are evaluated. The dataset and the scores are shown in the influential order, which can be referred for data cleansing.

Face Evaluation
Measure skin color of human face in input images, calculating a score called Individual Typology Angle (ITA) [2].

LIME (batch)
Using a method called LIME [3], the areas of the input image that affect the classification result are made visible in the model, which performs image classification. LIME(batch) plug-in processes all images in the specified dataset, while LIME plug-in processes a single image.

LIME (tabular)
Using a method called LIME [3], a classification result is explained with the contribution of the features in input table data. Each feature is explained with a set of inequality and degree of contribution, which enables to interpret the classifier judgement [4].

 

We will continue to update Neural Network Console.
We look forward to getting feedbacks from the users for futher improvements!

Neural Network Console Windows
https://dl.sony.com/app/

 

[1]
S.Hara, A. Nitanday, T. Maehara
“Data Cleansing for Models Trained with SGD” (2019)

[2]
Diversity in Faces
Michele Merler, Nalini Ratha, Rogerio S. Feris, John R. Smith
https://arxiv.org/abs/1901.10436

[3]
Why Should I Trust You ?: Explaining the Predictions of Any Classifier
Marco Tulio Ribeiro, Sameer Singh, Carlos Guestrin
https://arxiv.org/abs/1602.04938

[4]
Note that this plugin does not support regression model or classification model with categorical features.