Learning to Learn

Stochastic Depth — Training a Deep Network That's Sometimes Shallow Stochastic Depth — Training a Deep Network That's Sometimes Shallow Most regularization techniques you can name work at the level of activations or weights. Dropout zeroes out random neurons. Weight decay shrinks individual parameters. Label smoothing softens output distributions. These all share an assumption: the structure of the network is fixed, and we just perturb what flows through it. Stochastic depth makes a stranger move. It perturbs the network itself. Each training step, entire layers vanish. The 50-layer ResNet is sometimes a 48-layer network, sometimes a 45-layer one — the depth is a random variable. By the time training finishes, the optimizer hasn't really trained one network. It has trained an ensemble of subnetworks that share the same weights but differ in which blocks they actually use. This turns out to be a remarkably effective regularizer, and it is the main...

End-to-End Medical Image AI: From Raw DICOM to Clinical Deployment End-to-End Medical Image AI: From Raw DICOM to Clinical Deployment There is a question worth sitting with before writing a single line of code: when a radiologist reads a CT scan, they do not run three separate mental programs — one for loading the DICOM files, another for deciding which Hounsfield window to apply, a third for actually spotting the tumour. Something more practised happens. A trained visual system and decades of anatomical knowledge combine seamlessly into a single interpretive act. The pixels become a diagnosis. Building a machine that replicates even a fraction of that process requires solving a surprising number of sub-problems in the right order. This post walks through every stage of the pipeline — from where to get publicly available MRI and CT data, through preprocessing, model architecture, loss functions, training mechanics, and all the way to clinical deployment and ongoing mon...