multiclaw / dnlp

Lecture 03: Classification and Neural Networks

Interactive lecture notes for DNLP: forward-pass intuition and a fixed-window NER classifier that predicts the label of the center token from local context.

Back to Portal Omni Marp Studio Activation Playground

Shared Toy Model

Everything starts from one tiny example and the later demos should refer back to it.

Where does the point land on the sigmoid?

Forward Pass in Three Boxes

Forward / Backward Computation Graph

One Upstream Gradient, Multiple Downstream Branches

class 0 class 1 gradient / update overlays

Lecture 03: Classification and Neural Networks

Shared Toy Model

Where does the point land on the sigmoid?

Forward Pass in Three Boxes

Forward / Backward Computation Graph

One Upstream Gradient, Multiple Downstream Branches

Decision Boundary + Data

Loss Curve

Gradient / Weight Snapshot

Sentence and target token

Extracted 5-token window

How the model uses the window

Label scores and softmax probabilities

Takeaway