Explainable AI is the key to trustworthy AI, which ensures that AI decisions are transparent, accountable, and interpretable to decision makers. Modzy’s explainable AI framework produces explanations five times faster than open source methods like LIME and SHAP.

Explainability…Explained

Modzy’s explainable AI framework is inserted into an AI pipeline, enabling engineers, data scientists, and business leaders to see the explanation outputs alongside AI models within the Modzy platform.

So how does it work? The two images below demonstrate the solution in action.

An image of a cat was fed to Modzy’s Image Classification Model. The explainability output reveals what pixels the model is identifying within the image to determine it is a cat and not something else. This is illustrated by highlighting the pixels that are crucial to the identification process. The result? A correctly identified image with the reasoning for why layered over the image.

Data Input: Picture of a cat fed to an image classification model.

Data Output: Pixels highlighted to show why the model identified this image as a cat.

Leveraging adversarial attacks to understand model predictions

Modzy’s explainability solution leverages adversarial attack techniques allowing us to understand how a model behaves and how it reaches predictions.

By performing an adversarial attack on the image classification model, we can extract the attacks. Through extensive research and experimentation, the Modzy team devised a method to filter useful information from the attacks – a process where we extract crucial information with image segmentation to provide the final explanation of the model.

The below illustration breaks down the approach. For this example, Modzy’s explainability solution uses the cat image as the input of the image classifier, and the cat’s ears and nose are the key features for predictive analysis.

Figure 1(a) is the original cat image we are using as the input as we seek to understand why the model considers the image to be a cat and not something else. Next, Figure 1(b) represents an adversarial attack (on the cat image and and where the attack values are saved, Figure 1(c). The attack values are plotted as a histogram, Figure 1(b), and filtered by a threshold (the red lines). This threshold filters out useless and irrelevant features. Finally, the filtered attacks are mapped to the original, segmented image, Figure 1(d).) Figure 1(e) represents the segments within the image that were strongly attacked when trying to explain the model.

Modzy’s explainable AI framework provides advantages such as ease of interpretability and speed to insight that yield reduced infrastructure costs. Since output explanations are generated at the same time the model is performing inference, faster, more precise results are realized.

If you want to dive deeper into the technical aspects of Modzy’s explainability solution, check out “Explain This – Beyond Lime and SHAP: The Fastest Approach to AI Explainability”, a 30-minute Webinar from Modzy Data Scientist, Andrew Tseng.