Reproduce “COIN++: Neural Compression Across Modalities” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Emilien Dupont, Hrushikesh Loya, Milad Alizadeh, Adam Goliński, Yee Whye Teh, Arnaud Doucet
COIN++ compresses datapoints by representing them as coordinate-to-value functions and transmitting fitted modulations to a shared, meta-learned network, rather than a full network for each datapoint. On CIFAR10, its reported per-image encoding time was 94.9 ms, versus 29,700 ms for COIN. This comparison used different rate and reconstruction quality settings; the shared base network is not included in compressed size.
Key insights
- Employing only 'shift modulations' within SIREN's sine activation functions, mapped from a low-dimensional latent vector, provides superior compression rates and training stability.
- Meta-learning allows the base network to learn common data structures, drastically reducing encoding time to milliseconds by enabling individual data points to be encoded with just a few gradient steps.
- The learned modulations are highly robust to quantization, permitting significant bitwidth reduction (e.g., 32-bit to 5-bit) with minimal impact on reconstruction quality, unlike direct network weight quantization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.