This post walks through recent work on Discovering Latent Knowledge in Language Models Without Supervision by Burns et al. The paper uses latent knowledge in the model’s activations to train the model. Their method answers yes-no questions accurately by identifying a direction in the activation space that adheres to logical consistency properties, such as having opposite truth values for a statement and its negation.

Sigmoid functions are a type of mathematical function that has a characteristic “S” shape. They are commonly used in mathematical modeling to represent a variety of phenomena, such as the probability of an event occurring, the growth of a population, or the spread of a disease. They naturally exhibit the property of gradual then sudden increase without exploding. I use sigmoids all the time for fitting data. They are smooth and differentiable, as well as being easy to add boundary conditions to. In this post, I provide some tips for how to adapt them to different problem cases.

This post is going to explore SHAP values. SHAP stands for SHapley Additive exPlanations and is a way of explaining the output of a machine learning model.

In this post I want to talk about some techniques for dealing with skewed data, especially left-skewed data. Left-skewed data is a bit of a rarity. It’s something you don’t see very often, kind of like a left-handed unicorn. It can also be difficult to work with if you’re not prepared.

Local Interpretable Model-agnostic Explanations (LIME) is an important technique for explaining the predictions of machine learning models. It is called “model-agnostic” because it can be used to explain any machine learning model, regardless of the model’s architecture or how it was trained. The key to LIME is to “zoom in” on a decision boundary and learn an interpretable model around that specific area. Then we can see exactly how various factors affect the decision boundary. In this post, I’ll show how to use LIME to explain an image classification model.

This post is a tutorial demonstrating how to use Grad-CAM (Gradient-weighted Class Activation Mapping) for interpreting the output of a neural network. Grad-CAM is a visualization technique that highlights the regions a convolutional neural network (CNN) relied upon most to make predictions. While Grad-CAM is applicable to any CNN, it is predominantly employed with image classification models. This tutorial utilizes TensorFlow for implementation, but I made a parallel tutorial that works with PyTorch.

This post is a tutorial demonstrating how to use Grad-CAM (Gradient-weighted Class Activation Mapping) for interpreting the output of a neural network. Grad-CAM is a visualization technique that highlights the regions a convolutional neural network (CNN) relied upon most to make predictions. While Grad-CAM is applicable to any CNN, it is predominantly employed with image classification models. This tutorial utilizes PyTorch for implementation, but I made a parallel tutorial that works with TensorFlow.