Neural Decoding Using Generative BMI Models - Brain-Machine Interface Engineering - page 159

FIgURE 5.7: Stationary/moving classifiers.
forward model can be derived by the techniques discussed in Chapter 3 , specifically the gated com-
petitive experts. This allows multiple HMMs to map discrete portions of the neural data to complex
trajectories. Consequently, the individual forward models only learn a segment of the trajectory,
outperforming single forward model that must generalize over the full trajectory [ 41 ].
HMMs were utilized as gates in a gated competitive mixture of linear experts to differenti-
ate and model arm at rest from arm moving in a real-time scenario solely by analysis of the neural
recordings. Each one of these two possible outcomes are modeled by an HMM defined by param-
eters λ m (movement) and λ s (stationary), and the goal is to find which one is more likely given that
a sequence of neuronal firing rates is observed (Figure 5.7 ).
There were several difficulties that had to be conquered to apply HMMs to motor BMI data.
First, the input data are very high dimensional and discrete, that is 104 neurons of 100-msec binned
neural recordings. To decrease the number of parameters of an HMM that would model the poste-
rior density, we created a vector quantization (VQ) preprocessor using the Linde-Buzo-Gray (LBG)
algorithm [ 42 ] to decrease the training requirements for the HMM and achieve better results. A
second method to decrease the number of HMM parameters is to invoke the channel independence
assumption and implement an ensemble of single neural-channel HMM chains to form an indepen-
dently coupled hidden Markov model (ICHMM). Consequently, this classifier takes advantage of
the neural firing properties and removes the distortion associated with VQ while jointly improving
the classification performance and the subsequent linear prediction of the trajectory.
Vector Quantizing hMM. In this section, we broadly describe the VQ-HMM-based clas-
sifier illustrated in Figure 5.7 . We demonstrate an experiment in which arm movement is being