Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
OMiSO: Adaptive optimization of state-dependent brain stimulation to shape neural population states
Authors: Yuki Minai, Joana Soldado-Magraner, Byron M Yu, Matthew A. Smith
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We tested OMi SO using intracortical microstimulation with a Utah array and found that it outperformed competing methods that do not incorporate these advances. Taken together, OMi SO provides greater accuracy in achieving specified activity states, thereby advancing neuromodulation technologies for understanding the brain and for treating brain disorders. |
| Researcher Affiliation | Academia | 1Neuroscience Institute, Carnegie Mellon University 2Machine Learning Department, Carnegie Mellon University 3Center for the Neural Basis of Cognition 4Department of Electrical and Computer Engineering, Carnegie Mellon University 5Department of Biomedical Engineering, Carnegie Mellon University EMAIL |
| Pseudocode | No | OMi SO identifies an orthogonal transformation matrix ˆOi Rm m that fulfills: ˆOi = argmin O:OO =I Λ0(estable, :) Λi(estable, :)O 2 F (3) where estable is a list of nstable stable electrodes that are common to the reference session and the ith session, and F is the Frobenius Norm. This optimization can be solved in closed-form [17]. The ˆOi found is applied to Λi to obtain the aligned latent space Λi Rni m: Λi = Λi ˆO i (4) |
| Open Source Code | Yes | Python code for OMi SO is available on Git Hub at https://github.com/yuumii-san/OMi SO.git. |
| Open Datasets | No | We tested OMi SO using u Stim in a rhesus macaque monkey with a 96-electrode Utah array implanted in the PFC. |
| Dataset Splits | Yes | Of these sessions, 7 sessions were used for offline analyses ( Offline analysis sessions ), 10 sessions with randomly selected u Stim patterns were used to train the stimulation-response model and the stimulation-response inverse model ( Training sessions ), and 14 sessions were used to test the performance of different methods ( Test sessions ). |
| Hardware Specification | Yes | All models were implemented in Py Torch [26] and fit using the Adam W optimizer. We trained all models on a local computing cluster using 4 NVIDIA Ge Force RTX GPUs and 11GB of RAM. The same local computing cluster was used to run PPO-inspired adaptive model updates of the stimulation-response inverse model during u Stim experiments. Our experiments also involved three additional computers dedicated to specific tasks: one for experimental trial control (Intel Core i5-4590 @ 3.30 GHz 4, 15.5 GB RAM, Intel HD Graphics 4600, Ubuntu 20.04.2 LTS), one for neural data recording (Intel Core i7-7700 @ 3.60 GHz, 4 GB AMD Radeon R7 450 + Intel HD Graphics 630, Windows 10 Pro), and one for real-time neural data processing and u Stim pattern selection (11th Gen Intel Core i5-11500 @ 2.70 GHz 12, 15.4 GB RAM, AMD Radeon Pro WX 3200 + Intel Graphics, Ubuntu 20.04.4 LTS). |
| Software Dependencies | No | All models were implemented in Py Torch [26] and fit using the Adam W optimizer. We performed Bayesian Optimization (using Optuna [25]) with this error objective and selected the hyperparameter values that achieved the smallest error. |
| Experiment Setup | Yes | S1 Summary of hyperparameters The table below summarizes the hyperparameter values used in this study. The latent dimensionality m for the FA model was chosen by first computing the optimal dimensionality separately for each session by maximizing the cross-validated data likelihood, then finding the mode of the distribution of optimal dimensionalities across sessions. The number of electrodes used for latent space alignment was chosen to be smaller (by 1 to 8 electrodes) than the number of common usable electrodes to increase the robustness of alignment. For the stimulation-response model training, the learning rate, weight decay, and the number of training epochs (i.e., number of complete passes through the entire training dataset) were chosen based on a grid search. For the model inversion, we did not perform extensive tuning of the hyperparameters, such as a max epoch, since the model yielded similar performance (Eq. 11) even with different hyperparameters. Batch size per epoch refers to the number of synthetic data samples generated for each epoch. The parameters for u Stim pattern selection and the PPO-inspired adaptive model update were chosen by running simulations (Section S6). u Stim parameters (amplitude, frequency, and duration) were set to avoid causing overt behavioral changes or strong post-u Stim activity inhibition across the entire array. |