Teaching artificial intelligence to think like children may make them better learners. Kanishk Gandhi and Brenden Lake at New York University found that a common assumption children make while learning, called the mutual exclusivity bias, would make AIs better at learning tasks such as processing language.
Mutual exclusivity bias refers to the incorrect assumption that once an object has a name or a label, it can鈥檛 also have another. 鈥淎 child will find it difficult to call a poodle a dog if it knows it鈥檚 a poodle,鈥� says Gandhi.
However, for early language learners like toddlers the bias is useful. When they come across a new word it is more likely that it will correspond to a genuinely new object or meaning, so making this assumption allows them to learn new things faster.
Advertisement
To see whether AIs show the same learning bias, Gandhi and Lake trained 400 different neural networks 鈥� a type of AI algorithm 鈥� to associate pairs of words and their meanings. The AIs were then tested on ten previously unseen words.
The AIs more regularly predicted that a new word would correspond to a known meaning rather than an unfamiliar one, suggesting that they don鈥檛 have exclusivity bias.
The team then looked at datasets used to train AIs to translate between languages, which consisted of pairs of sentences in two languages, to see if an AI would benefit from exclusivity bias. For example, assuming that a new word in a source language (e.g. xylophone in English) would be translated into a new word in the target language (e.g. xylophon in German), as opposed to a familiar word.
Based on probability, the team predicted that for a large part of the early training, the AI was better off assuming that unfamiliar words are more likely to correspond to unfamiliar words in a second language. This suggests it would be advantageous to bake the bias in to the AIs. 鈥淢utual exclusivity will facilitate algorithms to perform better,鈥� says Ghandi.
As children learn more, the bias slowly wanes as the likelihood that we鈥檝e never encountered something before decreases 鈥� and AIs could do the same. Similar to humans, when an AI becomes more knowledgeable in a particular area, the bias could decrease over time, says Gandhi.
搁别蹿别谤别苍肠别:听arXiv,听
