Exploring Neural Networks for Recognition of Hand-drawn Shapes
summary
A working iPad demo recognizes circles squares and triangles by feeding 28×28 bitmap inputs into a simple feed‑forward network (one hidden layer, ~280 nodes) implemented via a Swift library, using grayscale pixels, Bezier-path stroke normalization and backpropagation training. The system trains on-device by saving user sketches and typically converges after about 50–100 examples; it tolerates noisy strokes but struggles with unseen rotations and requires a confidence threshold for rejection. Core ML is avoided because training happens on the device.
AI-written summary · about
chapters (33)
- 0:00Intro
- 1:13Overview
- 2:28Motivation
- 8:09Neurons
- 10:58Sigmoid
- 11:56Graphing
- 14:09Feedback
- 14:47Example
- 17:00Outputs
- 17:45Learning
- 18:43Mean Square Error
- 20:14Gradient Descent
- 21:20Complicated Gradient Descent
- 23:06Tensorflow Demo
- 27:01Prototype
- 27:37Diagram
- 28:18Features
- 30:28Parameters
- 32:12Demo
- 35:19Building the model
- 35:46Least squared error
- 36:19The library
- 37:15Why not CoreML
- 38:28Make Your Own Neural Network
- 39:15Interface
- 40:51Update
- 42:21Capture
- 43:02Path
- 43:48Smooth Path
- 44:05Preparing the Bitmap
- 46:24Training
- 47:02Rejection
- 48:12Presentation Materials
related talks
40mBuilding a Handwriting Recognition System
34mThe Potential of IMAGE Processing in Swift for Developers
57mWriting a Neural Net from Scratch
21mRecognizing Human features using Deep Networks
37mJavaScript Machine Learning from Magic to Science
55mCore ML: Vision
33mComputer Vision, Not Just for Breakfast Anymore
44mAI for every developer