Jumping knowledge networks let each node combine embeddings from every layer, so the effective neighborhood size can vary across the graph. I explain the influence distribution argument and the LSTM attention aggregator. Then I apply a jumping knowledge GAT to cortical segmentation, where it holds up better than a plain GAT as depth increases.
Wang et al. argue that graph attention networks overfit their attention weights and oversmooth signals across class boundaries. Their fix adds two margin-based losses and aggregates only over the top-k attended neighbors. I walk through the method and my implementation of the layer in DGL.
A loss term that penalizes predicting cortical labels that aren’t adjacent to the true label. Implemented in PyTorch and DGL and used as a regularizer alongside standard cross-entropy.
Implementing a graph convolution layer whose filters are learned Gaussian kernels over node features, written as a custom DGL message-passing layer for segmenting the cortical surface.
Principal angles between two subspaces come from the SVD of the product of their orthonormal bases. Most subspace distances, including the Grassmann distance, are functions of those angles.
Notes from a lab meeting talk on how to structure, build, and upload a Python package to PyPI, with an example from my own code.
Wrapping a Python script in a bash script so it can be submitted to a Sun Grid Engine cluster with qsub.
Computing a gradient map on a cortical surface mesh by regressing scalar values onto neighbors projected into the tangent plane. The gradient map is then segmented with a priority-flooding watershed algorithm built on a heap-based priority queue.
A friend asked how to tell whether a biomarker can predict cardiotoxicity. I build ROC curves by hand on synthetic case and control data, then use them to assess the diagnostic power of simulated dose-response curves.
Deriving the delta method from a Taylor expansion and simulating its use for variance stabilizing transformations of Poisson and exponential data and for standard errors in polynomial regression.