Neural Noise- A software platform supports real-time and adaptive neuroscience experiments, including closed-loop BCI and neural signal processing pipelines.1
- The platform is directly useful for neural signal processing and experimental design in neural interfaces.1 1
Weekly enrichment (2026-07-20)
- The platform is improv (Nature Communications, 2025, s41467-025-64856-3), a modular Python system for constructing and orchestrating adaptive, closed-loop neuroscience experiments where models drive ongoing data collection and experimental manipulations in real time.2
- improv is built on a simplified “actor model” of concurrent computation: each pipeline stage is a user-defined actor running in its own process, communicating by message passing, with large data (e.g. images) held in a shared in-memory store built on the Plasma library from Apache Arrow to minimize copying.2
- In an in-silico benchmark on two-photon calcium imaging of larval zebrafish (GCaMP6s), improv ran CaImAn Online segmentation and fit a linear-nonlinear-Poisson model online via stochastic gradient descent over sliding windows, with the online fit converging to the offline estimate after a single repetition of the stimulus set.2
- For closed-loop stimulus optimization, an improv Bayesian-optimization actor used a Gaussian-process model with an upper-confidence-bound acquisition function to find neuronal direction tuning: on average just ~15 stimulus presentations characterized peak tunings across 300 neurons, and the online peak matched the offline fit for 93% of neurons, versus a 576-stimulus grid that would take ~200 h for the whole brain.2
- improv drove real-time all-optical experiments in zebrafish expressing the red-shifted opsin rsChRmine, autonomously selecting direction-selective neurons for two-photon photostimulation (1045 nm) during 920 nm GCaMP6s imaging, typically stimulating ~25 target neurons per session with 5 repetitions.2
- A latent-dynamics demo used proSVD streaming dimensionality reduction plus the Bubblewrap model on monkey motor-cortex spiking (182 units reduced to 6 dimensions); predictions one full second (~100 samples) ahead dropped only 11% in log predictive probability versus one-step-ahead.2
- The authors position improv against related tools — BRAND (C++, Linux-only), Heron (Windows), Bonsai (C#/Windows), and Cleo (biophysical closed-loop simulation) — emphasizing cross-platform, domain-agnostic Python operation validated on hardware from an 8 GB MacBook to ~$4k desktops; the code is open source at github.com/project-improv/improv.2
- Such real-time model-in-the-loop infrastructure is directly applicable to closed-loop BCI and neurofeedback, where decoding and intervention must occur within tight latency budgets; other frameworks in this space include CLEF, Falcon, and BRAND.3