Abstract:
Inductive generalization is adaptive in novel contexts for both biological and artificial intelligence. Spontaneous generalization in inexperienced animals raises questions on whether predispositions (evolutionarily acquired biases, or priors) enable generalization from sparse data, without reinforcement. We exposed neonate chicks to an artificial social partner of a specific color, and then looked at generalization on the red-yellow or blue-green ranges. Generalization was inconsistent with an unbiased model. Biases included asymmetrical generalization gradients, some preferences for unfamiliar stimuli, different speed of learning, faster learning for colors infrequent in the natural spectrum. Generalization was consistent with a Bayesian model that incorporates predispositions as initial preferences and treats the learning process as an update of predispositions. Newborn chicks are evolutionarily prepared for generalization, via biases independent from experience, reinforcement, or supervision. To solve the problem of induction, biological and artificial intelligence can use biases tuned to infrequent stimuli, such as the red and blue colors.
摘要:
归纳泛化在生物和人工智能的新环境中是自适应的。在没有经验的动物中自发的泛化引发了关于易感性(进化获得性偏见,或先验)从稀疏数据中实现泛化,没有加固。我们将新生小鸡暴露于特定颜色的人工社交伙伴,然后查看了红黄色或蓝绿色范围的概括。泛化与无偏模型不一致。偏见包括不对称的泛化梯度,一些对不熟悉刺激的偏好,不同的学习速度,更快地学习自然光谱中罕见的颜色。概括与贝叶斯模型一致,该模型将偏好作为初始偏好,并将学习过程视为偏好的更新。新生小鸡在进化上已经为推广做好了准备,通过独立于经验的偏见,钢筋,或监督。为了解决归纳法的问题,生物和人工智能可以使用针对不常见刺激的偏见,如红色和蓝色。