In a small model, sleep (stop training and run noise so each connection updates by looking only at the two neurons it joins) was confirmed to improve performance. But no evidence was found that the same holds in large models.
What formatting and flashing erase first is not the contents of the files but the map that records what is where (the partition table and the start of the file system), so data that has not been overwritten is still there and can be recovered. By erasing the map with a flash, changing how much the new image overwrites and trying to bring everything back, and watching how File Explorer, Disk Management and the status of each file change, we see why. The first page on computer storage.
A film is one view of what a CT scan records, so a virtual film computed from the CT can be matched to it. By turning and moving a 2D toy CT to match a film, and by switching between the cone model and the parallel model and changing the source distance, we see why the cone-beam model is what makes registration work. The first page on X-ray and CT imaging.
One question tells pre-training, post-training and inference apart: do the weights change? Following the smallest GPT, trained by us, step by step, we see that the two kinds of training change the weights while inference changes only the context, and how reinforcement learning changes the output without a right answer, by drawing and scoring. The second page on language models.
Kimi K3.1 Pro has the same base model as K3, yet its AA index rose from 44 to 55. Most of the gain came from tests of using tools to finish a job, and that strength came from much larger post-training (reinforcement learning). The page also shows how reinforcement learning changes a model with scores alone, without right answers. The first page on language models.
A microinjector pushes the sample from behind the needle with air (pneumatic) or with oil (hydraulic). Air compresses easily and oil hardly at all, so a pneumatic injector sets the pressure and the time, and a hydraulic one sets the volume. Change the needle tip and the sample, and see how much each delivers and what pressure it takes. The second page on robotics and automation.
Robot types are divided by how the joints are connected, and the structure sets where the hand can reach and the hand’s pose. Change the hand’s pose and see whether Cartesian, SCARA, articulated, delta, parallel-link arm and Stewart platform robots can do five lab jobs. The first page on robotics and automation.
Even the neurons of the middle layer, which have no answer, get their own share of the error when how wrong the output is passes back. Understand it by stepping through how a small multilayer perceptron learns XOR. Compare an MLP with a single-layer perceptron, the sigmoid with the step function, and gradient descent walking down the slope, in one figure.
train-val-test is not legacy: calib is one more split for the values fitted with data after training, the temperature T and the threshold. Change the split that fits the two values and move the threshold yourself. The first page on machine learning.
How a phase contrast microscope makes colorless, transparent cells, which brightfield can hardly tell from the background, visible with two rings, the condenser annulus and the phase plate inside the objective: set the annulus number, aperture, centering and focus yourself. The first page on microscope imaging.
How YOLOv3 spreads nine anchors over three grids (13 × 13, 26 × 26, 52 × 52) for large and small objects, unfolds deep features and concatenates them with earlier ones, and teaches only one anchor per object that it holds an object: drag objects and stack the 107 layers into a pyramid. It follows the YOLOv2 page.
How YOLOv2 shares objects among five anchor boxes per cell and folds fine 26 × 26 features into the 13 × 13 grid with a passthrough layer: drag objects, pick anchors, and fold a feature map. It follows the YOLOv1 page.
IOU measures how much two boxes overlap, and NMS keeps one of the boxes that overlap on an object: move the boxes and step through the algorithm. Background for the YOLOv1 page.
How YOLOv1 splits an image into a 7×7 grid and finds every object in one pass of the network, explored through the architecture and outputs of a model built from the paper.