A neural network written in jq (2017)
9 hours ago
- #jq
- #MNIST
- #neural-network
- A neural network implemented entirely in jq processes the MNIST dataset with a 94% success rate.
- The program uses a JSON configuration and processes data in an immutable manner, making state saving/loading straightforward.
- Input records include 'input', 'expected', and 'train' fields, with backpropagation performed when 'train' is true.
- Progress is logged via debug statements, and the final output includes error rate and counts of processed records.
- The neural network library is in 'neural_net.jq', with sample configuration in 'config-trivial.json'.