An added SEEG model gave an uncertain loss gain in 16 later sessions at C=0.01
We tested recall forecasts in 16 public SEEG session pairs from OpenNeuro ds004809 version 2.2.0. Each person supplied 180 test words. We combined fixed constant, task, and task plus SEEG models. Each word position had separate weights. We changed weights only after a recall list ended. At C=0.01, the added-SEEG mixture had mean loss 0.600755 nats per word. The two-model task mixture had loss 0.604190. The paired gain was +0.003436, with a 95% person bootstrap interval of [-0.001511, +0.009595]. Nine of 16 people had a positive gain. At C=0.1, the added model increased mean loss by 0.013470. A diagnostic task-copy control also left the primary gain uncertain. All 64 tests met the known Bayesian bound of 12 log(K) excess loss over the best fixed expert. The bound assumes unchanged models and recall labels before the next list. Features become available 1.3 seconds after word onset. This exploratory replay reuses previously studied people. It does not measure stimulation effects. The package preserves code, features, labels, four figures, fitted models, source records, and retrieval instructions. A full offline replay matched the saved results. Fresh checks passed for 96 raw windows from all 32 sessions.
stereoelectroencephalography · SEEG · recall forecasting · online learning · log loss · late labels · Bayesian mixtures · memory · brain-computer interface
References
4- usesAdded SEEG power increased recall forecast loss in all eight later sessions with a fixed logistic reader sci:sha256:5d98ad3a4b35585f15bfec73be164b2de5da77c70912a6e3c5282178fbe7d256
The source reader study supplied the fixed power-feature setup and the first eight session pairs carried into the direct parent package. We used its prior transfer-loss evidence to frame the late-label mixture question.
- usesA fixed SEEG recall reader with scaling from 60 early words had higher mean loss than task inputs in eight new session pairs at C = 0.01 sci:sha256:cc2e373832a1a405c2992d004e962cfdf49dcfddb914d6b2ba0dbdb2bf23ff97
This package supplied all 16 selected SEEG session pairs, power features, word labels, source metadata, pinned raw ranges, raw witnesses, and fixed reference model parameters. We used its task inputs and calibration objective.
- usesdoi:10.18112/openneuro.ds004809.v2.2.0
Source dataset: public bipolar SEEG recordings from OpenNeuro ds004809 version 2.2.0.
- doi:10.1038/s41467-017-02753-0
Cited by
2 publicationsPublic catalog entries on Sciscape.
- usesOne fitted neural weight per session reduced mean log loss against a fixed weight in 16 reused SEEG session pairs at C=0.01
This work supplied the 16 session pairs, word features, event tables, source pins, extraction jobs, and MIT source-fit helpers used in this study. All people and test words are reused.
- usesFeedback with assumed correct labels at recall end lowered mean log loss and AUC in 16 reused SEEG session pairs at C=0.01
Used its benchmark selection, 96 power features, task inputs, source-fit objective, and pinned raw retrieval recipe through the direct parent.