Home

So from the previous post, we took a brief look over the pixel attribution approaches. This time, we’ll go through another linage of interpretability - concept-based approaches. Let’s start from what ‘concept’ is, and why we need it. Again, this post follows the contents from here. TCAV: Testing with Concept Activation Vectors Pitfalls of feat...

(Can we skip to the good part) Yes, after a while, I decided to skip all those ‘classical’ stuffs of interpretable ML, and proceed to neural network analysis! I know, you and I all love neural networks, and thanks to Christoph, their NN part is exclusive to web so we can peek and learn from the extensive writings. I’ve been extremely busy with g...

For some reason, I’m now working on the sampling methods to improve the models’ performance. We have tons of data, but using every row is pretty expensive and shows slower convergence. After some research, I found an insightful reference of training data sampling for neural networks. Although it’s not clear if this is applicable to my project, l...

Long time no see - it has been a month since I started my work in Quora. I started my ML Engineer position at Distribution team, which is highly related to the content recommendations. Speaking of recomendation, I would like to introduce the two-tower model1, which is a part of the (almost) industry-standard architecture for any recommendation s...

Overview After all those interpretable models, we now go forward to model-agnostic methods. One obvious merit among many pros is that model-agnostic methods do not depend on the model type we’re interested in. Ribeiro et al.1 claims that there should be three desirable aspects of a model-agnostic explanation system: Model flexibility: indepe...

RuleFit1 is the extension of decision rules, which can be easily summarised: classifying instances by several IF-THEN statements. We will skip the introduction for the decision rules, so please refer Molnar’s page if you’re interested in the model. Decision rules technique has their own distinct strategies to enhance interpretability and accurac...

Logistic regression and GLM are somehow direct extension from the linear regression so I skip the post about those. Instead, in this post, we will take a look for another simple yet powerful method, decision tree. After we go through the basic of the decision tree, we will also take a look for the paper named “Neural-Backed Decision Trees”, whi...

Here I skip the definitions and some qualitative concerns for the interpretability and move on to the interpretable models from Molnar’s reference1. The models introduced in this post are more like statistically founded approaches, which means they might not be able to applicable to complex tasks such as speech recognition or object detection. ...