Brain-Machine Interfaces (BMI) have strong potential to benefit a large number of disabled people but current decoding algorithms suffer from the following shortcomings. First, BMI decoding algorithms are often trained offline, but this paradigm ignores the discrepancy between the Manual Control (MC) and the Brain Control (BC) modes of operation. Second, the standard neural decoder, the Kalman filter, does not explicitly take into account the control of movements by neural activity. To address these problems, we propose a biologically motivated neural decoder structure by explicitly adding a control signal and unmeasureable neural activity. Since the parameter estimation problem is underdetermined, we propose a new parameter estimation method that minimizes the discrepancy between the MC and BC. We demonstrate the effectiveness of our methods by synthesizing MC and BC data in a Linear Quadratic (LQ) optimal control setting with a partial loss of neural control in BC, and show that the proposed decoder is more robust to a partial loss of neural control than a standard Kalman filter that does not utilize any reparameterizations.