Finding Biomarkers to Predict Seizures
Finding Biomarkers to Predict Seizures
Embedded Files
This ongoing project uses BrainBERT, a model designed to analyze brain activity, to identify patterns that could predict seizures before they happen. By applying mechanistic interpretability, the goal is to look inside BrainBERT to uncover these patterns, leading to more transparent and effective tools for predicting seizures.Â
This ongoing project uses BrainBERT, a model designed to analyze brain activity, to identify patterns that could predict seizures before they happen. By applying mechanistic interpretability, the goal is to look inside BrainBERT to uncover these patterns, leading to more transparent and effective tools for predicting seizures.Â
Part 1: Fine-tuning a Foundation Model to Detect Seizures
Part 1: Fine-tuning a Foundation Model to Detect Seizures
This project involves fine-tuning BrainBERT, a Transformer foundational model, to detect seizures in stereo-EEG data.
Part 2: Using Sparse Autoencoders to Look Inside the Model
Part 2: Using Sparse Autoencoders to Look Inside the Model
This project uses sparse autoencoders and interpretability methods to to distill BrainBERT embeddings into key features.
Bias in Breast Cancer Detection
Bias in Breast Cancer Detection
This project introduces a novel approach to identifying biases in the Emory EMBED dataset by analyzing how the model processes information. Instead of manipulating input data, we used Grad-CAM to visualize and interpret the ResNet model's behavior in identifying breast cancer, offering unique insights into both the model's decision-making process and potential data biases.
This project introduces a novel approach to identifying biases in the Emory EMBED dataset by analyzing how the model processes information. Instead of manipulating input data, we used Grad-CAM to visualize and interpret the ResNet model's behavior in identifying breast cancer, offering unique insights into both the model's decision-making process and potential data biases.
Part 1: Generating Grad-CAM Heatmaps
Part 1: Generating Grad-CAM Heatmaps
Grad-CAM heatmaps highlight the regions in the image that are most important for the model's decision
Part 2: Heatmaps Reveal ResNet Bias
Part 2: Heatmaps Reveal ResNet Bias
Heatmaps reveal that the model mistakenly focuses on irrelevant text in image corners rather than breast tissue when detecting cancer.
Page updated
Google Sites
Report abuse