Abstracts / Neuroscience Research 71S (2011) e108–e415 e201 P2-t11 Analysis of classifier variability between sessions and subjects for constructing brain machine interface using multiple subject’s data Okito Yamashita , Masa-aki Sato Brain Information Communication Research Laboratory Group, ATR, Kyoto, Japan The brain machine interface (BMI), which employs brain signals to directly control devices such as a computer and a robot, collects much attention both in neuroscience and engineering communities. BMI requires a subject- specific translator from brain signals to control commands. For constructing such a translator, a calibration experiment and classification-rule learning are necessary. In a calibration experiment, brain activities while a subject carries out repetitions of two or more kinds of tasks are recorded. Then a classification rule of brain activity patterns between the task conditions is learned. A real-time experiment can be conducted only after constructing the subject-specific translator. Successful control of devices in the real-time experiment relies on accuracy of the constructed translator. However a num- ber of data recorded in a calibration session is not sometimes sufficient to learn an accurate translator. One approach for solving this problem is to take advantage of data from previous sessions (different subjects or the same subject in past). Fazli et al. (2009) have proposed a method to construct a subject-independent classifier from a data bank consisting of previously recorded data. Alamgir et al. (2010) have proposed the Bayesian frame- work to combine one subject’s data with another subjects’ data. However there is no study that investigates when the previous sessions data improve the current subject data. Here we conduct classification analysis of thirty EEG experimental sessions during which three subjects conducted motor imagery tasks. We will demonstrate conditions that result in successful use of previous session data. References Fazli, et al., 2009. Neural Network 22 (9), 1305–12. Alamgir, M., et al., 2010. 4th International BCI Meeting 4 (H-2), 1. doi:10.1016/j.neures.2011.07.868 P2-t12 Reconstruction of movement-related intracortical potentials from micro-electrocorticogram signals in mon- key primary motor cortex Hidenori Watanabe 1 , Masa-aki Sato 2 , Takafumi Suzuki 3 , Mitsuo Kawato 4 , Yukio Nishimura 1,5,6 , Atsushi Nambu 6,7 , Tadashi Isa 1,6 1 Div of Behav Dev, Natl Inst for Physiol Sci, Okazaki, Japan 2 ATR Neural Infor- mation Analysis Lab, Kyoto, Japan 3 Grad Sch of Information Sci and Technol, Univ of Tokyo, Tokyo, Japan 4 ATR Computational Neuroscience Lab, Kyoto, Japan 5 PRESTO, JST, Tokyo, Japan 6 Grad Univ for Advanced Studies, Hayama, Japan 7 Div of System Neurophyiol, Natl Inst Physiol Sci, Okazaki, Japan Decoding the elctrocorticogram (ECoG) recorded with subdural electrode array can provide better spatial and spectral resolution of cortical activities than scalp electroencephalography. The intracortical local field potentials (LFPs) in the motor cortex, which are the electric field caused by transmem- brane currents of cortical neurons flowing near the electrode, would carry substantial information about the behavioral outcome. LFP recording, how- ever, includes the risks attendant upon penetration of the brain parenchyma. Reconstructed LFP from ECoG is expected to provide useful signal for a brain–machine interface (BMI). Here, we examined the reconstruction of movement-related LFP from ECoG. Subdual 32 ch micro-ECoG (regular grid arrangement with 1 mm inter-electrode distance) and 64 ch-LFPs were simultaneously recorded from primary motor cortex, while a monkey per- formed constrained reaching and grasping to capture foods moving in the 3D free-workspace. Trajectories of the arm movements were detected by infrared optical motion capture system. LFP profiles located in various depths could be predicted by sparse logistic regression (r = 0.71 ± 0.02 with 23-fold cross validation at 0.2 mm from cortical surface). Furthermore, the power of beta (10–35 Hz) and high-gamma (100–150 Hz) frequency signals of the reconstructed LFP varied depending on the status of the arm movement as well as those of the actual LFP signal (r = 0.80 ± 0.06, r = 0.48 ± 0.06 between the reconstructed and actual LFPs at 0.2 mm from cortical surface at beta and high-gamma frequency range, respectively). Our results indicated successful reconstruction of motor-related LFPs from ECoG by sparse logistic regression. Research fund: Strategic Research Program for Brain Sciences from the MEXT, Japan. doi:10.1016/j.neures.2011.07.869 P2-t13 Metal pin electrode for brain–machine interface Kouji Takano 1 , Shigeru Toyama 3 , Tomoaki Komatsu 1 , Yasoichi Nakajima 2 , Kenji Kansaku 1 1 Sys Neurosci Sect, Dept Rehab Brain Fnct, Res Inst of Natl Rehab Center, Tokorozawa, Japan 2 Dept Rehab Brain Fnct, Res Inst of Natl Rehab Center, Tokorozawa, Japan 3 Biotech Sect, Dept Rehab Eng, Res Inst of Natl Rehab Center, Tokorozawa, Japan Most of the typical electrodes for EEG-based brain–machine interface (BMI) have been using conductive pastes to reduce the impedance between scalp and electrodes. To use electrodes with paste requires not only elaborate work for preparation but also for removal of paste sticking on hair after measure- ment. Here, we developed a system with paste-less metal pin electrodes for BMI. The system consists of 7 metal pin electrodes and sensor mounting ele- ment with spring for each pin electrode. The system can provide adequate force to maintain sensor contact with scalp. The metal pin electrodes have rough surface, so that it can reduce contact resistance between scalp and electrodes. We also prepared a pre-amplifier for each system. By using the system, we successfully obtained brain waves, and the signals were almost equivalent to those observed with the conventional paste-based electrodes (less than 20 k). Moreover, it does not require not only elaborate work for preparation but also for removal of paste sticking on hair after measurement. Further development of the metal pin electrode is now in progress. doi:10.1016/j.neures.2011.07.870 P2-t14 Canonical correlation analysis (CCA) of multi-joint motion (JM) and dorsal root ganglion (DRG) neuronal activ- ities Jun Morimoto 1 , Tatsuya Umeda 2 , Yukio Nishimura 2 , Tadashi Isa 2 , Mitsuo Kawato 1 , Keisuke Toyama 1 1 BICR, ATR, Kyoto, Japan 2 Dept. Dev. Physiol., Natl. Inst. Physiol. Sci The present neuroscience is in need of the means for ensemble-wise cor- relation (EC) analysis to analyze massive data sampled by simultaneous multi-unit recording. We resolved this issue by introducing the canonical correlation analysis (CCA) to estimate EC between JM and DRG neuronal activities. We tested the CCA performance by simulation in two ways. First, CCA was conducted on DRG neuronal activities modeled as the products between sinusoidal inputs to the JM receptors and the JM–DRG connectivity matrix (CM) containing significant divergence and convergence components, and the JM–DRG CM was reconstructed from the EC determined by the CCA. The CCA performance evaluated as the coincidence of the reconstruct with the real one was high (0.95). Next we estimated the forward CM from the JM receptors to the DRG and the inverse CM from the DRG back to the JM recep- tors, and reconstructed DRG activities and JM from the model JM and DRG activities, respectively. The CCA performance evaluated as the variance of the reconstructs from the real ones was also rather small (error rate = 0.13 and 0.30). Thus the simulation study guaranteed the CCA performance. We con- ducted CCA on the DRG neuronal activities (n = 32) record by multi-channel recording during passive multi-joint (N = 7) motion of the upper limb in the anesthetized monkey. The CCA reconstructs of the forward and inverse CM were in fine agreement (coincidence = 0.77). The CCA revealed that the JM receptor-DRG circuitry is not parallel straight lines as has been supposed, but conveys significant divergence and convergence that realize step-wise integration of motor information across joints from the elbow to the finger tip as the motor information modules and sparse representation of those mod- ules in the DRG neuronal ensemble, respectively. This study shows that the CCA has the power to reveal both structural and informational integration in the neural circuitry. Research fund: SRBPS, MEXT. doi:10.1016/j.neures.2011.07.871 P2-t15 Decoding brain activity with smooth sparse regres- sion Matthew de Brecht 1 , Noriko Yamagishi 1,2,3 1 NICT, Brain ICT Lab., Kyoto, Japan 2 ATR-CMC, Kyoto, Japan 3 JST-PRESTO, Saitama, Japan Sparse regression has been shown to be a useful method for decoding high- dimensional fMRI and MEG data by automatically selecting relevant feature dimensions. However, when applied to signals with high spatio-temporal correlations, sparse regression often over-prunes the feature space, which can result in overfitting and weight vectors that are difficult to interpret. To help overcome this problem, we propose a modification of L1-normed sparse