Google Unveils Machine Learning Model for AI-Driven Image Classification

2019-10-15

It is commonly believed that as the complexity of AI systems increases, their interpretability decreases. Researchers have started to address this issue using libraries like Facebook's Captum, which explains how neural networks make decisions using deep learning frameworks such as PyTorch, IBM's AI Explainability 360 toolkit, and Microsoft's InterpretML. To make AI decisions more transparent, teams from Google and Stanford have recently studied a machine learning model called Automatic Concept Explanation (ACE), which can automatically extract meaningful visual concepts.


As explained by the researchers in their paper, most machine learning explanation methods alter individual features (e.g., pixels, superpixels, word vectors) to approximate the importance of each feature to the target model. This approach is flawed because even the smallest input values can be easily manipulated. In contrast, ACE identifies higher-level concepts by training a classifier and using a set of images as input before extracting concepts and determining their importance. Specifically, ACE segments images at multiple resolutions to obtain object textures, object parts, and object levels. It then groups similar fragments into examples of the same concept and outputs the most important parts.


To test the robustness of ACE, the team used Google's Inception-V3 image classifier model, which was trained on the ImageNet dataset, and selected a subset of 100 categories out of the dataset's 1,000 categories to apply ACE. They observed that concepts marked as important tend to align with human intuition, such as the badge on a police car being more important than the asphalt on the ground. However, this is not always the case. In images of basketball, the most important concept is the player's jersey rather than the basketball itself. The researchers acknowledge that ACE is far from perfect and struggles to effectively extract exceptionally complex concepts. However, they believe that the insights it provides into the model's learning associations could promote safer use of machine learning.


The researchers write in their paper, "We experimentally validate the significance and consistency of the model and further confirm that they do carry signals for prediction. Our method of automatically grouping input features into higher-level concepts is highly practical. It is crucial for correctly predicting elements present in an image when these meaningful concepts appear as coherent examples."

RECENT AI NEWS

RECENT AI TOOLS