Volitional Generation of Reproducible, Efficient Temporal Patterns

Abstract

One of the extraordinary characteristics of the biological brain is the low energy expense it requires to implement a variety of biological functions and intelligence as compared to the modern artificial intelligence (AI). Spike-based energy-efficient temporal codes have long been suggested as a contributor for the brain to run on low energy expense. Despite this code having been largely reported in the sensory cortex, whether this code can be implemented in other brain areas to serve broader functions and how it evolves throughout learning have remained unaddressed. In this study, we designed a novel brain-machine interface (BMI) paradigm. Two macaques could volitionally generate reproducible energy-efficient temporal patterns in the primary motor cortex (M1) by learning the BMI paradigm. Moreover, most neurons that were not directly assigned to control the BMI did not boost their excitability, and they demonstrated an overall energy-efficient manner in performing the task. Over the course of learning, we found that the firing rates and temporal precision of selected neurons co-evolved to generate the energy-efficient temporal patterns, suggesting that a cohesive rather than dissociable processing underlies the refinement of energy-efficient temporal patterns.

Keywords: brain–machine interfaces; energy-efficient code; precise temporal patterns; primary motor cortex.

Figure 1

Schematic of precise temporal codes. (A) Schematic illustrating two types of precise temporal codes. Shaded rectangles mark the precise synchronization beyond rate fluctuation that is preserved in the spike trains of two neurons. However, the efficient precise temporal pattern depicted on the right contains less spikes than that on the left, although the number of spiking coincidences under the two cases is identical. (B) Schematic of how efficient codes are studied on artificial neural networks by explicitly incorporating the regularization term into the cost function in order to penalize high firing rates $\mathbf{r}\left(t\right)$. Here, $\mathbf{c}\left(t\right)$ denotes the target signals to be represented by the network. (C) Schematic of the explicit incorporation of information capacity $\mathbf{c}\left(\mathbf{t}\right)$ and metabolic cost (neurons’ firing rates $\mathbf{r}\left(\mathbf{t}\right)$) in the decoder of the brain–machine interface for the study of the efficient codes in vivo. $NCS\left(t\right)$ is the decoded variable defined in this study.

Figure 2

The BMI-based task paradigm. (A) The task was run in a closed-loop manner. Neural activity from the subject’s M1 was automatically read and extracted. All leads and lags between spikes from the trigger and target units were used to compute the NCS, which was further fed back by mapping it to the frequency of an audio cursor and the vertical position of a visual cursor. In this exemplified trial, the NCS threshold for reward was 0.32. The threshold for the water reward was set based on the NCS distribution estimated in the first session. (B) Task structure of one typical session. The session started with a 5-min baseline block, after which the distribution of NCS could be estimated. One trial could span 15 s at most, followed by a 4 s long inter-trial interval (ITI). The NCS was computed using neural activity in a 300 ms sliding window.

Figure 3

Proficiency increased in performing the BMI-based task. (A) The success rate exhibited a growing trend over the course of learning (linear fitting test to determine whether slope was significantly greater than zero: B11: p = 0.005; C05: p = 0.008). (B) The trial duration was reduced over the course of learning (linear fitting to account for sample size N and scatter among replicates and to test whether slope was significantly greater than zero: B11: p = 0.048; C05: p = 0.010). (C) Pronounced increase in the success rate from the early phase to the late phase, pooling over two subjects. One-tailed Mann–Whitney test, early < late, p < 0.0001. (D) Significant shortening in trial duration from the early phase to the late phase, pooling over two subjects and all trials (Mean ± SEM), F(1,1729) = 12.78, p = 0.0004. Interaction between two factors: F(7,1729) = 1.14, p = 0.3338. *** p < 0.001, **** p < 0.0001.

Figure 4

The stronger modulation of NCS in the last session of learning indicates that reproducible neural patterns were generated. (A,B) The NCS distribution of the task block in the first session (S1) compared to the last session (S10). (C,D) Distribution on the right side of the panel was shown in the y-axis, logarithmically scaled to better reveal the growing frequency of high NCS, especially the range over the NCS threshold (i.e., 0.3–0.5), as indicated by the red box.

Figure 5

Rewarding neural patterns of BMI-based tasks were temporally precise. (A) CCHs of two subjects Attn AE - part of Figure 5 caption in the last session showed pronounced coincidence between two units in a time window spanning from 0 to +15 ms. Error bars denote the SD rendered from the resampling analysis on jittering. The asterisks above the bars indicate a significant coincidence. **** p <0.0001. (B) The CCHs of two subjects in the last session with higher resolution showed finer temporal granularity. Error bars denote the SD rendered from the resampling analysis on jittering. The asterisks above the bars indicate a significant coincidence. * p < 0.05, *** p < 0.001, **** p < 0.0001.

Figure 6

Efficient firing in modulating NCS for the BMI task. (A) The spike counts (averaged over trigger unit and target unit) of the rewarding neural patterns from the task block were significantly lower than those of the sampled neural patterns from the baseline block (Mack–Skilling test, B11: p < 1 × 10${}^{-4}$; C05: p < 1 × 10${}^{-4}$). The spike counts were grouped into five blocks according to the coincidence scores of the corresponding neural patterns. The black line represents the median. (B) Average modulation index after pooling over all significantly modulated indirect neurons across sessions (mean ± SEM). The asterisks above the lines indicate a significant positive departure from zero, whereas those below represent values smaller than zero. * p < 0.05.

Figure 7

Evolution of rewarded neural patterns throughout learning. (A) The CCH coincidence in [0, 15] ms is correlated with the firing rates of the target unit for both subjects (Left: B11, p = 1.05 × 10${}^{-4}$; Right: C05, p = 9 × 10${}^{-4}$). However, the evolution showed opposite directions, as indicated by the arrows. The coordinates of the last session are labeled. (B) Rewarded neural patterns of both subjects under the same neural space. Each dot represents the averaged neural pattern in one session. Neural patterns from stable sessions are indicated by the orange dots.