Seeking a research collaborator — compact recurrent architecture for edge/industrial signal tasks
Mehdi Zarrouky
Hi all,
I'm an independent researcher working on a compact recurrent neural network architecture aimed at extremely resource-constrained settings (sub-kilobyte models, microcontroller-class deployment). The design is built around provable spectral stability and contractive dynamics, which makes it unusually robust under INT8 quantization and very parameter-efficient.
Across several public benchmarks (bearing fault diagnosis, remaining-useful-life estimation, human activity recognition from IMU data, keyword spotting), the architecture reaches accuracy comparable to standard GRU/LSTM/CNN baselines while using roughly one order of magnitude fewer parameters, and it degrades far less under post-training quantization than typical autoencoder baselines. I've also documented several cases where it does not help — I'm interested in honest characterization, not a sales pitch.
I'm looking for a collaborator with a strong background in one or more of: recurrent/state-space models (S4, Mamba, LRU and related), dynamical systems / Koopman theory, or embedded ML and signal processing. Someone who would enjoy stress-testing the theory, finding where it breaks, and thinking about which application domains it genuinely fits.
I should be transparent about two things. First, my own background is not deeply technical — the architecture was developed through a long iterative process, and I'm aware I need a real expert to pressure-test and extend the theoretical side. Second, some implementation details are not yet public for IP reasons, so an initial conversation would be under a mutual NDA — but the benchmarks, the evaluation protocol, and the theoretical framing are all open for scrutiny.
This is an open-ended collaboration proposal, not a paid position — I'm exploring whether there's a basis for working together, including co-authorship and, if it makes sense down the line, a more formal partnership. I'd rather be upfront about that than waste anyone's time.
If this resonates, send me a message and I'm happy to share more, including a private technical brief once we've established a basic confidentiality agreement.
Thanks for reading.
Mehdi Zarrouky
zarr...@gmail.com