清华大学脑与智能实验室研究团队于2026年10月26日在成都和清华大学发布最新研究成果，探讨了人工智能并不必然需要无限增加算力的可能性 1。该团队通过实验证明，当神经网络仅保留约5%的连接时，其表现依然优于全连接神经网络 1。这一发现挑战了“人工智能越聪明就需要越多算力”的传统认知，为构建更高效的AI模型提供了新的技术路径 1。
研究者卡洛·维托里奥·坎尼斯特拉奇提出了名为“稀疏连接”的关键技术，该方法允许人工智能系统自行调整内部信息传递的路径以提升效率 1。坎尼斯特拉奇将这一机制比喻为协作模式的优化：“不是让所有人一直开会，而是让真正需要协作的人找到彼此” 1。尽管当前大多数计算设备并非专门为这种能够灵活改变连接方式的网络架构而设计，但这项研究揭示了未来更高效的人工智能可能摆脱对远端服务器的依赖，转而具备在本地设备上运行的潜力 1。此外，针对机器人等物理实体应用，若单纯依靠无限堆砌算力来提升性能，将不可避免地导致耗电量大增、散热困难以及重量增加等问题，进而缩短设备的续航时间 1。
A research team from the Brain and Intelligence Laboratory at Tsinghua University has challenged the assumption that artificial intelligence efficiency requires unlimited increases in computational power. 1 Led by researcher Carlo Vittorio Canistraci, the team proposed a "sparse connectivity" technology that allows AI systems to adjust their internal information transmission pathways to improve efficiency. 1 Experiments showed that neural networks retaining only about 5% of their connections still outperformed fully connected networks. 1
Canistraci explained the concept with an analogy: "It is not about keeping everyone in constant meetings, but rather letting those who truly need to collaborate find each other." 1 The study noted that current computing devices are not specifically designed for networks that flexibly change connection methods. 1 Additionally, researchers pointed out that if robots indiscriminately increase computing power, it would lead to higher energy consumption, heat dissipation issues, increased weight, and shorter battery life. 1 This research suggests that future highly efficient AI could potentially run on local devices instead of relying solely on remote servers. 1
评论
还没有评论，欢迎留下第一条。