Latest posts
-
A Kalman Update From GPS And Motion
An IMU can carry a position estimate between GPS fixes, but its errors accumulate. A one-dimensional worked example shows how a Kalman update combines an uncertain prediction with a noisier-or-better measurement, and where the assumptions enter.
-
What Cross-Entropy Teaches a Classifier
A classifier’s raw scores are not probabilities, and cross-entropy does more than reward the right answer. A small three-class example shows how logits become a loss and what its gradient asks the model to change.
-
From Camera Pixels to a 3D Ray
A pixel coordinate is not a direction until you account for focal length, the principal point, and the camera’s coordinate frame. Work through the conversion, then see where lens distortion changes the result.