Not compressing neural networks themselves, but using them to compress other things. This means using neural nets to reconstruct signals (images, audio, video) with low error from a small (in bits) summary, especially alongside existing noisy data transmission pipelines. Maybe we could even try both at once and think about minimum description length. That might be interesting.
- Descript Audio Codec / descriptinc/descript-audio-codec: “State-of-the-art audio codec with 90x compression factor. Supports 44.1kHz, 24kHz, and 16kHz mono/stereo audio.” (Kumar et al. 2023)
- Image Compression with Neural Networks – Google AI Blog
- mlomnitz/DiffJPEG
- rshin/differentiable-jpeg: Code for “JPEG-resistant Adversarial Images”
1 With vector embeddings
Thanks to the success of vector embeddings we can frequently represent in a weird vector space which looks suggestively like classic encodings. In NLP, the embeddings are usually larger than the original text, though, so can we actually compress this way?
Maybe. There have been a bunch of works in that domain recently (Duggal et al. 2024; Miwa et al. 2025; Yan et al. 2025). The most viral one is (Bachmann et al. 2025), which has an elegant demonstration.
