資訊協同對視網膜預測能力的貢獻

為了避免危險或捕捉獵物，動物必須預測以彌補視覺處理途徑的延遲。先前的研究表明視網膜神經節細胞(RGC)具有對移動物體的預測信息編碼的能力。為了更系統性地研究視網膜的預測能力，實驗中，我們使用隨機移動的光條或全域均勻變化的光強度刺激牛蛙視網膜以及多電極系統測量視網膜神經節細胞的電位，並使用互信息(Mutual Information)分析視網膜的激發頻率(firing rate, $\gamma(t)$)與不同時間的刺激($\xi(t+\delta t)$)的關係。我們更進一步的利用部分信息分解(PID)分解$\gamma$與$\xi$和$\dot\xi$̇的聯合隨機變數的互信息，並發現$\gamma(t)$關於$\xi$和$\dot\xi$的協同資訊(synergy)於自$\gamma$的編碼形式為$\xi$和$\dot\xi$的線性組合，且視網膜能夠基於$\xi$和$\dot\xi$之間的協同資訊生成預測。另一方面，我們通過來自白噪聲刺激與發放-觸發平均方法(STA)測量視網膜的核(線性算子，$K$)。並且我們嘗試用$K$再現視網膜對於有和沒有空間相依性的兩種不同刺激的預測行為。結果顯示視網膜中對均勻光強度的預測可以發生視網膜前端並且可以線性描述，而對移動物體的預測則不能用類似的線性機制描述。

關鍵字

視網膜；預測；隨機過程；互訊息；部分信息分解

並列摘要

To avoid dangers or catch prey, animals must make predictions to compensate for delays in their visual processing pathway. Predictive motion information is encoded in the spiking activities of retinal ganglion cells. To systematically study the predictive properties of a retina, stimuli in the form of a stochastic moving bar or stochastic whole-field light intensity are used in experiments with a bullfrog retina in a multi-electrode system. Herein, the anticipatory encoding was analyzed in the retina by measuring the mutual information (MI) between the firing rate $\gamma(t)$ and stimulus $\xi(t+\delta t)$ with a time lag. Furthermore, the partial information decomposition of the MI between $\gamma$ and the joint variable of $\xi$ and $\dot\xi$ was investigated and the retina was found to be capable of anticipation based on the synergistic information generated between $\xi$ and $\dot \xi$ Moreover, the synergy stemming from the encoding form of $\gamma(t)$ which was a linear combination of $\xi$ and $\dot\xi$. Our results suggest that illusions such as the anticipation studied here during retinal perception can originate from the recombination of information extracted in the retinal network. The kernel $K$ was also measured, by the spike-triggered average from white noise stimulation. Additionally, we tried to reproduce the predictive behavior of the retina using $K$ in two kinds of stimulations, with and without spatial dependency. The results show that the prediction of homogeneous light intensity in the retina occurs in the first synapse and can be linearly described; however, the prediction of a moving object cannot be described in a similar linear mechanism.

並列關鍵字

Retina ； Anticipatory Dynamics ； Stochastic Process ； Mutual information ； Partial Information Decomposition

參考文獻

D. Balduzzi and G. Tononi. Integrated information in discrete dynamical systems: motivation and theoretical framework. PLoS computational biology, 4(6):e1000091, 2008.