A neural network written completely in jq (2017) (github.com)

A surprising demo implements a working neural network entirely in jq (the JSON query language), using a small C helper (make_json.c) to convert MNIST into JSON and then piping records through jq for training and testing. The example reportedly reaches about 94% accuracy on MNIST (error_rate ≈ 0.0593) and emits final stats as JSON. While not competitive for speed—runs can take days—the project is a proof‑of‑concept showing that a declarative, JSON‑centric tool can express feedforward and backpropagation logic. Technically, the implementation treats the network state as the accumulator in a reduction over input records because jq’s data is immutable. Each record has fields "input", "expected", and a "train" boolean; train=true performs backpropagation, false just updates the error count. Configuration is passed as a JSON argument (example config-trivial.json), the network code lives in neural_net.jq, and the runner is example/example.jq. Outputs and progress are JSON/debug lines (e.g., total/train/test/errors), and the final JSON contains error_rate, num_errors, and counts. The project highlights expressive power, portability (state easily saved/loaded as JSON), and educational value, while also illustrating practical limits of using an interpreter designed for data transformation rather than numeric performance.