Cognitive psychologists often use the term $\textit{fluid intelligence}$ to describe the ability of humans to solve novel tasks without any prior training. In contrast to humans, deep neural networks can perform cognitive tasks only after extensive (pre-)training with a large number of relevant examples. Motivated by fluid intelligence research in the cognitive sciences, we built a benchmark task which we call sequence consistency evaluation (SCE) that can be used to address this gap. Solving the SCE task requires the ability to extract simple rules from sequences, a basic computation that in humans, is required for solving various intelligence tests. We tested $\textit{untrained}$ (naive) deep learning models in the SCE task. Specifically, we tested two networks that can learn latent relations, Relation Networks (RN) and Contrastive Predictive Coding (CPC). We found that the latter, which imposes a causal structure on the latent relations performs better. We then show that naive few-shot learning of sequences can be successfully used for anomaly detection in two different tasks, visual and auditory, without any prior training.
翻译:认知心理学家经常使用美元(textit{fluid 智能)这一术语来描述人类在没有任何事先培训的情况下完成新任务的能力。 与人类相比,深神经网络只有在经过大量相关实例的广泛(预)培训之后才能执行认知任务。 我们借助认知科学的流体智能研究,建立了一个基准任务,我们称之为序列一致性评估(SCE),可用于弥补这一差距。 解决 SEC任务要求有能力从序列中提取简单的规则,这是解决各种情报测试所需的人类基本计算。 我们在 SCE 任务中测试了$(textit{untrated} (naviductive) $(na) 深学习模式。 具体地说,我们测试了两个可以学习潜在关系的网络,即关系网络(RN) 和对比性预测聚合(CPC) 。 我们发现后两个网络对潜在关系的因果关系结构表现更好。 我们然后显示,对序列的天真少的学习可以成功地用于在两种不同任务(视觉和审计)中的异常检测。