Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Curiosity-driven networks develop brain-like modular structures in study

AI summary

A study led by Francesco Poli of the University of Cambridge combined eye-tracking data from 275 infants aged 8 to 15 months with artificial neural networks trained on 20 tasks. The most curious networks generalized better to unseen tests and developed interconnected, modular, brain-like structures. Their connection density first rose and then fell, a pattern the researchers compared with synaptic overgrowth and pruning in human development. By comparing connectivity curves with human developmental curves, the team estimated that the most curious, brain-like networks corresponded to about age 20, while networks that both lacked curiosity and had an unsuitable wiring pattern corresponded to about age 9. Alireza Modirshanechi of the University of Göttingen called the structural finding surprising but said it requires verification, including whether non-curious networks would develop similar structures with much longer training. The report cites a bioRxiv preprint.

In-depth insight

The experiments indicate that combining intrinsic curiosity with constraints that penalize costly long-range wiring can produce modular, brain-like network organization. This offers a possible route to flexible, human-like AI alongside approaches based on increasing model scale, although the structural result still requires verification.

IT之家 11 月 2 日消息,人类大脑为何演化得如此复杂?答案或许可以归结为好奇心。研究人员训练了数千套学习模式各不相同的人工神经网络,他们发现,好奇心最强的网络不仅拥有更复杂、灵活的认知能力,还会自发形成类似人脑的结构。

这些网络会形成相互连通的模块,并且遵循和人类相近的发育规律:比如先长出大量密集神经连接,再逐步修剪精简。剑桥大学的弗朗切斯科 · 波利(Francesco Poli)表示,好奇心通常被视作复杂大脑带来的结果。“但有没有可能,好奇心反过来塑造了大脑的形态?”

为了模拟婴儿的学习方式,波利和同事采集了 275 名 8 至 15 个月大婴儿的眼动追踪数据。研究人员在屏幕上播放一系列彩色图形,图形序列会提供不同信息量,暗示下一个图形出现的位置。波利说:“位置不是随机的,每一组图形都有一套可供学习的规律。”

#Francesco Poli#University of Cambridge#Alireza Modirshanechi#University of Göttingen#bioRxiv#ChatGPT

Original · 5 min readRead the original
Group chat

No comments yet. Start the conversation.