HN 提问：神经形态计算会取代传统人工智能吗？
3 分•作者: lennart-rth•2 个月前
我最近在思考当前深度学习的核心低效问题时，了解到神经形态计算。
如今的 AI 严重依赖于超密集层、所有层和神经元持续不断的计算，以及全局反向传播来寻找最优权重更新。但当我们审视人脑时，这些情况都不存在。
人脑的运作原理与现代大型语言模型 (LLM) 截然不同：
* **局部演化：** 神经元在很大程度上是独立的，它们根据局部邻域和简单的反馈回路（如神经递质，例如多巴胺）进行演化，而不是基于全局误差信号。
* **极端稀疏性：** 该系统具有极高的稀疏性（神经元仅在参与了激活链后才会演化和更新）。
* **事件驱动处理：** 神经元仅在被评估并实际被其他脉冲神经元触发时才会发放脉冲。
相比之下，我感觉当前的 LLM 似乎是通过同时训练所有层和神经元，并尝试寻找最优的全局更新函数来强行解决问题。
如果我们审视过去几年 AI 发展的轨迹，会发现一个清晰的模式：
* 2020–2022 年：扩大数据集和原始计算规模。
* 2023–2024 年：扩展上下文窗口并转向专家混合模型 (MoE)。
* 2024–2025 年：思维链和推理时推理。
* 2025 年至今：自主执行和并行多智能体系统。
从根本上说，这些进步中的每一个都只是扩展计算和处理令牌的不同方式。因此，这种扩展自然会在某个时候遇到瓶颈，因为电力是有限的。
这就是为什么我相信长期的进步不可能仅仅来自于永无止境的扩展。它需要的是根本性的效率提升。那么，神经形态计算是否能实现这一点，还是它尚未成熟？
查看原文
I’ve recently learned about neuromorphic computing while thinking about the core inefficiencies of current deep learning.<p>Today AI relies heavily on super-dense layers, continuous computation of all layers and neurons at all times, and global backpropagation to find optimal weight updates. But when you look at the human brain, none of that happens.<p>The brain operates on principles that stand in contrast to modern LLMs:<p>- Local Evolution: Neurons are largely independent, evolving based on their local neighborhood and simple feedback loops like neurotransmitters (e.g., dopamine) rather than a global error signal.
- Extreme Sparsity: The system is massively sparse (neurons only evolve and get updated when they have been involved in a firing-chain).
- Event-Driven Processing: Neurons are only firing when they get evaluated and are actually triggered by other spiking neurons.<p>In contrast, current LLM‘s seem to me like they brute-force their way through problems by training all layers and neurons at the same time and trying to find the optimal global update function.<p>If you look at the trajectory of AI advancements over the last few years, a clear pattern emerges:<p>- 2020–2022: Scaling up datasets and raw compute.
- 2023–2024: Expanding context windows and shifting to Mixture of Experts (MoE).
- 2024–2025: Chain-of-thought and inference-time reasoning.
- 2025–Present: Autonomous execution and parallel multi-agent systems.<p>Fundamentally, every single one of these advancements is just a different way of scaling up compute and processed tokens.
So this scaling will naturally hit a wall at some point as electricity is not unlimited.
That’s why I believe that long term progress can not come from just scaling forever. What it needs is radical efficiency improvements. So could neuromorphic processing be exactly that or is it not mature yet?