Intrinsic Variable Learning for Brain-Machine Interface Control by Human Anterior Intraparietal Cortex

Abstract

Although animal studies provided significant insights in understanding the neural basis of learning and adaptation, they often cannot dissociate between different learning mechanisms due to the lack of verbal communication. To overcome this limitation, we examined the mechanisms of learning and its limits in a human intracortical brain-machine interface (BMI) paradigm. A tetraplegic participant controlled a 2D computer cursor by modulating single-neuron activity in the anterior intraparietal area (AIP). By perturbing the neuron-to-movement mapping, the participant learned to modulate the activity of the recorded neurons to solve the perturbations by adopting a target re-aiming strategy. However, when no cognitive strategies were adequate to produce correct responses, AIP failed to adapt to perturbations. These findings suggest that learning is constrained by the pre-existing neuronal structure, although it is possible that AIP needs more training time to learn to generate novel activity patterns when cognitive re-adaptation fails to solve the perturbations.

Keywords: anterior intraparietal cortex; brain-machine interface; individual-neuron learning; intrinsic-variable learning; posterior parietal cortex; spinal cord injury.

Figure 1.

(A) A graphical representation of the BMI learning paradimg with a single neuron. (B) The stimulus-response rules for the BMI-calibration (left), the BMI-pro task (middle) and BMI-anti task (right). Each panel illustrates the firing rate (MEAN ± SEM) of a hypothetical trained neuron for successful trials in each of the three tasks.

Figure 2.

The firing-rate distributions of two different single trained-neurons for each of the two stimuli in the BMI-pro and BMI-anti task blocks in (A) session 1 and (B) session 3 for both correct and incorrect trials.

Figure 3.

(A) The temporal dynamics of the firing rates for a trained AIP neuron that directly contributed to the BMI output in a typical session. (B) Similar to panel A, but for two untrained AIP neurons. (C) The percentage of untrained neurons that flip their activity between BMI-pro and BMI-anti tasks after the participant adopted the anti-wrist movement strategy.

Figure 4.

(A) A graphical representation of the BMI learning paradigm with two neurons. (B) A simplified, conceptual illustration of the stimulus-response rules for the BMI-pro and BMI-fsb trials for the pair of stimuli 1 and 5. Points correspond to the hypothetical firing rates of the two trained neurons in response to four stimuli (color coded), in the BMI-calibration.

Figure 5.

The firing rates for the two trained neurons in response to the stimuli 1 and 4 (colored coded) in (A) the 8-target BMI-calibration, (B) BMI-pro, and (C) BMI-fsb tasks. The black solid and dashed lines represent the linear thresholds used for decoding in the BMI-pro and BMI-fsb tasks, respectively. (D) The firing rates of the same two trained neurons in response to targets 3 (cyan) and 7 (red) in the BMI-calibration (the arrows indicate the relationship between the stimulus and matched target locations). The two stimuli (cyan and red) correspond to the matched target locations.

Figure 6.

(A) The proportion of trials for the eight targets that the participant verbally reported to have attempted a wrist-movement in response to stimuli 1 and 4, during the BMI-fsb task in the example session described in Figure 5. (B) Correspondence matrix for the BMI-fsb session presented in Figure 5. It shows the proportion of trials for the eight targets decoded by the trained vs. untrained neurons. The x axis represents the target decoded by the trained neurons and the y axis the target decoded by the untrained neurons in the example session described in Figure 5. (C) Probability distribution (blue trace) across 15 sessions of the difference between the targets decoded by the trained and untrained neurons in the BMI-fsb task when the target decoded from the trained neurons was a matching target. The red trace describes the probability distribution from 15 sessions of the difference between the targets decoded from the trained and untrained neurons in the 8-target BMI-calibration when the target decoded by the trained neurons was the BMI-calibration target. The peak at zero corresponds to the case in which the trained and the untrained neurons encode the same matching target. Error bars represent SEM and red and blue lines show the spline interpolation to the data.

Figure 7.

(A) The firing rates for the 2 trained neurons in response to the eight stimuli (color/shape coded) in the 8-target BMI-calibration. The pair of stimuli 2 and 7 was selected in the BMI-pro and BMI-nfsb tasks. The black solid and dashed lines represent the linear boundaries used for decoding in the BMI-pro and BMI-nfsb tasks, respectively. (B) Percentage of trials that fall into the left (red) and right (blue) side of the BMI-nfsb task threshold in response to the eight stimuli in the 8-target BMI-calibration. (C) The firing rates for the two trained neurons in response to the stimuli 2 and 7 in the BMI-pro. To be successful, the activity of the two trained neurons should fall below and above the BMI-pro threshold in response to stimuli 2 and 7, respectively. (D) Similar to C, but for the BMI-nfsb task. To be successfully, the activity of the two trained neurons should fall on the right and left side of the BMI-nfsb threshold in response to stimuli 2 and 7, respectively (E) Proportion of successful trials (i.e., performance) across 6 sessions. Successful trials were achieved when the activity of the trained neurons during the go period fell on the correct side of the boundary. Participant achieved performance accuracy 94.03% ± 2.65% (MEAN ± SEM) in the BMI-pro and 87.22% ± 3.0791 in the BMI-nfsd across 6 sessions (no significant difference in the performance; two-tailed t test, p=0.1324) for the cognitively solvable stimulus – i.e., stimulus that could be acquired by adopting a cognitive strategy. However, the performance was significantly deteriorated from 88.90% ± 3.93% in the BMI-pro to 50.0% ± 5.55% in the BMI-nfsb (two-tailed t test p < 0.001) across the 6 sessions for the cognitively non-solvable stimulus – i.e., stimulus that could not be acquired by adopting any cognitive strategy. Instead, an individual-neuron mechanism was required to generate novel patterns of activity and was not accomplished during each daily session.