Experimental study of Bayesian inference based on experience

本研究采用一种改进的延迟样本匹配（DMTS）任务来研究动物（鸽子）的贝叶斯推理，在该任务中，在比较刺激呈现之前，先呈现哪一侧是正确的信息刺激信号。在此过程中，样本呈现频率在1:9和3:7两种情况下变化，信息刺激的信度也在100%到80%之间变化。在匹配性能稳定后，引入两个不同的测试环节，其中一个条件下不提供样本，而另一个条件下既不提供样本也不提供信息刺激。测试结果表明，在所有被试中，频率较低的刺激的选择比例与不提供样本时DMTS任务中使用的两个样本的比例大致一致，而在不提供样本和信息刺激时，频率较低的刺激的选择比例与贝叶斯定理的预测一致。实验结果表明，在DMTS任务中，动物受试者在学习用比较匹配样本时，能够检测到两个样本所占比例的差异，并对信息刺激的可靠性进行评估。虽然目前的研究提供了基于经验的贝叶斯推理的数据理解，但更多的研究应该检查基于经验的贝叶斯推理的比较和发展方面，以及对程序的可能修改。

The present study investigated bayesian inference in animals (pigeons) by using a modified delayed matching-to-sample (DMTS) task in which information stimulus signaling which side is correct was presented prior to the presentation of comparison stimuli. In this procedure, sample presentation frequency was varied in two conditions (I.e., 1 : 9 and 3 : 7 conditions) and reliability of the information stimulus was also varied from 100% to 80%. After matching performance was stabilized, two different test sessions were introduced, in which no sample was presented in one condition, whereas neither sample nor information stimulus was presented in another condition. The results from testing showed that in all subjects, choice proportions for the less frequent stimulus were approximately consistent with the proportion of two samples used in the DMTS task when no sample was presented, and further that choice proportions for the less frequent stimulus were consistent with predictions derived from Bayesian theorem when neither sample nor information stimulus was presented. These results revealed that animal subjects can detect the difference in the proportion of two samples and evaluate the reliability of information stimulus when they learn to match samples with comparisons in the DMTS task. Although the present study provided data understanding Bayesian inference based on experience, additional studies should examine comparative and developmental aspects of Bayesian inference based on experience as well as possible modifications of the procedure.