Title : A closed-loop active inference framework for robust neural signal decoding in clinical BCI

Neural decoding of visual stimuli from non-invasive electroencephalography (EEG) remains a critical bottleneck in Brain-Computer Interface (BCI) research, driven by low Peak Signal-to-Noise Ratio (PSNR) and the unpredictable, changing nature of continuous brain activity. Standard feed-forward deep learning architectures treat decoding as a passive mapping task, failing to adapt to the temporal drift and inter-subject variability inherent in human EEG recordings. In this study, we propose a Closed-Loop Active Inference (CLAI) framework that integrates principles of predictive coding with Generative Adversarial Networks (GANs) to achieve robust neural decoding.

The proposed framework consists of a U-Net-based generator coupled with a semantic “Oracle” – a pre-trained critic network representing known clinical state structures. Unlike traditional methods, CLAI performs test-time latent space optimization: rather than relying on a static forward pass, the model iteratively refines its internal belief state z through gradient-based feedback from the Oracle, minimizing structural reconstruction error. We validated the proposed framework on the UCI Epileptic Seizure Reconstruction dataset, a high-noise classical benchmark. Our empirical results show that CLAI achieves a significant +6.56dB improvement in PSNR over open-loop baselines. Furthermore, visual audits demonstrate that the active inference loop facilitates the recovery of critical geometric features related to seizure activity – features that remain obscured in purely feed-forward architectures. By shifting from passive inference to active refinement, this work provides a scalable, structurally aware paradigm for clinical neural decoding and real-time brain-state monitoring.

Sai Pradyumna Thiriveedi is an undergraduate researcher at the Sri Sathya Sai Institute of Higher Learning, pursuing a B.S. in Mathematical Sciences and Computing (Computer Science major), and Rank 1 course topper for three consecutive semesters. He is a Summer Research Fellow (also IAS–INSA–NASI SRFP 2026 for Amritha University) at IIT Hyderabad's MOS Lab, working on physics-informed deep learning for digital holographic microscopy. His research spans generative modeling, explainable AI, and biomedical signal/image analysis, including EEG-to-image neural reconstruction and automated retinal diagnostics. He is also President of CodeEX, SSSIHL's programming and product engineering club.