Convolutional Neural Networks in APL
summary
Implements a MNIST/LeNet-style CNN in APL using ten one-line APL building blocks that avoid explicit indexing, showing convolution, pooling and fully connected layers as composable array operators; proves index-free convolution by shifting, multiplying and summing subarrays; runs training for comparison (TensorFlow init ~6s, interpreted APL training much slower, a compiled array-language version narrows the gap); identifies missing pieces: GPU support and automatic differentiation for production-grade performance.
AI-written summary · about
chapters (7)
- 0:00Limitations of ML frameworks
- 3:07Problem definition
- 5:21Neural network architecture
- 8:35Implementing operations
- 17:15Scaling and model design
- 22:10Performance analysis
- 23:38Future directions
related talks
27mAPL-Inspired Techniques for Advanced NumPy
58mFunctional Array Funhouse Intensive - Part 2 of 5
01h 07mParallel-by-construction Tree Manipulation with APL (Part 2)
44mMachine Learning with Artificial Neural Networks in Clojure
55mA Tour of Deep Learning With C++
21mConvolutional Neural Networks at scale in
30mMachine Learning & Neural Networks Demystified
02h 50mFunctional Array Funhouse Intensive - Part 3 of 5