neurOS-v1 Package: neuros-neurofm — the neuroFMx Foundation Model

Building a Multimodal Neural Foundation Model

neuros-neurofm is the architectural centerpiece: neuroFMx, a foundation model built from scratch to treat heterogeneous neural signals as tokens and learn shared representations across modalities, subjects, and sessions.

What It Comprises

• Tokenizers — per-modality tokenizers for spikes, LFP, EEG, fMRI, calcium, audio, video, and binned rates, with a temporal_alignment module to put them on a common clock.
• Backbone — a Mamba state-space backbone for linear-time modeling of long neural sequences, plus a popt population transformer.
• Models & heads — neurofmx, neurofmx_complete, neurofmx_multitask, and multimodal_neurofmx, with single- and multitask heads.
• Objectives — masked-modeling and multitask losses for self-supervised pretraining.
• Adapters — LoRA and unit-ID adapters for parameter-efficient cross-session / cross-subject transfer.
• Deployment — model_compression, a neuros_adapter integration, and multimodal training.

How It Differs

neuros-foundation wraps published models; neuros-neurofm is my own model. Its separable tokenizer, alignment, backbone, adapter, and loss layers are where "combining and building large multimodal models" becomes concrete, recomposable engineering rather than a slogan.

Part of the neurOS-v1 platform.