CHAPTER 4 ESTIMATION OF KINEMATIC-DRIVEN NEURAL REPRESENTATIONS FOR
4.7 D ISCUSSION
This study proposed a new approach for finding kinematics dependent neural representations of M1 population activity. We estimated kinematics dependent neural representations by nonlinearly mapping M1 population firing rates to kinematics dependent latent factors (KLDF), where the nonlinear mapper was approximated by DCCA. We validated the proposed neural representations by examining the neural dynamics information presented in the neural trajectory of them and by decoding neural representations into various kinematic parameters. We compared KLDF with other neural representations obtained by SCC, FA, GPFA, and LFADS. We found that neural trajectories produced by KDLF revealed neural dynamics before and after movement onset more clearly than those by other methods (Figs.4-3 ― 4-5) and that decoding KDLF enabled a simple linear decoder to predict kinematic parameters more accurately than decoding other neural presentations (Figs. 4-6 ― 4-9). Notably, KDLF showed consistently high decoding performance for different monkeys performing dissimilar tasks.
To validate the decoding performance of the proposed approach, we predicted four kinematic parameters from KDLF, including velocity, acceleration, jerk, and speed, and compared decoding performance with those with four other neural representations. Decoding KDLF yielded higher performance for predicting velocity and acceleration compared to all other counterparts and slightly higher performance for predicting speed than LFADS that was the second-best model. Note that this study did not attempt to investigate the effect of using multiple stitched sessions, in which LFADS has Figure 4-9 Comparison of reconstruction errors for position. Open circle denotes average of the mean squared errors between actual and reconstructed position for tested trials, and each error bar indicates the standard error. Each color corresponds to monkey C, M, and F, respectively. Asterisks denote significant difference between neural representations (** p < 0.01, a two-way ANOVA by a post hoc analysis with a Tukey-Kramer correction for multiple comparison
74
shown advantages for maintaining decoding performance. Nevertheless, the decoding result in the present study suggests that the proposed neural representations of KDLF would improve motor decoding for intracortical BMIs. Note that LFADS does not support real-time generation of neural representations yet, whereas KDLF can be generated and decoded in real time once the training of DCCA is completed. A further study will investigate a feasibility to apply our method for stitched sessions to achieve stable online BMI performance.
According to [25], neural trajectories of a complex kinematic model present clear rotational patterns for behavioral events (e.g., a reach toward a specific target in a single trial). Based on the output-null model that describes how neural activity drives target muscles [35], we observed that the neural trajectory produced by KDLF remained constant before motion onset with a small variability across target directions, reflecting neural states in a null space, and that neural trajectories began to dynamically vary by sweeping the jPC space during arm movement after motion onset. These characteristics of neural trajectories associated with arm movements were most pronounced with KDLF, compared to other neural representations. Thus, we speculate that more reliable and clear representation of movements in KDLF might be linked to improved decoding.
Neural representations are related to generating muscle activation patterns involved in complex movements [23], [26], [35]. For instance, muscles that generate a specific movement would be driven by a linear combination of neural activities [35]. However, since it is difficult to predict complex kinematic information from noisy neural manifolds with a simple linear model, a nonlinear model such as deep neural networks (DNNs) has often been proposed [9], [71], [163]. In this study, we also utilized DCCA to approximate nonlinear mapping between neuronal population activity and KDLF that is supposed to represent various kinematic information. Although it is possible to use other models than DCCA for our approach, we adopted DCCA due to its several advantages. As an alternative approach to DCCA, we considered a multilayer perceptron (MLP) and compared its performance with DCCA on the prediction of KDLF. We observed that DCCA could yield significantly higher performance than MLP in terms of the reconstruction error (7.6 % higher on average). In addition, it has been demonstrated that DCCA can effectively improve velocity prediction accuracy by decoding canonical variates of DCCA and be used for extracting features from a large-scale dataset [37], [149].
It is noteworthy that the proposed neural representations appeared to provide movement-related information better when the arm movements became more complex: e.g., from 2D tasks (monkeys C and M) to 3D task (monkey F). This was observed in both neural trajectories and decoding performance.
We speculate that neural representations directly related to kinematic parameters (i.e., KDLF) might embed kinematic information more clearly and thus become more effective than other unsupervised neural representations for encoding more complex movements. Note that KDLF suggests a new way of dealing with the same latent factors of M1 population activity that are also processed by Gaussian process (GPFA) and dynamical systems (LFADS), with a more potential to improve motor decoding.
75
This study proposed a novel neural representation of neuronal population activity via supervised learning based on kinematic information. It thus seems natural to obtain results that neural representations found by supervised learning yielded better performance of decoding kinematics than those by unsupervised learning without any kinematic information. So, it is reasonable to consider that previous neural representations found in unsupervised ways are more useful for investigating intrinsic nature of neuronal population activity whereas our neural representations are more focused on extracting specific kinematics information. Yet, our approach may provide a better way to generate useful neural representations applicable to BMI.
76