机器学习研究者们，让我们攻克硬件工程难题
1 分•作者: MRiabov•6 个月前
大家好，HN 的朋友们，我在这里与机器学习研究人员交流。
目前，机器学习领域存在一个极具挑战性但尚未被充分探索的问题——硬件工程。
到目前为止，所有因素似乎都对解决这个问题不利——预训练数据基本不存在（不像 NLP/计算机视觉那样丰富），该领域的研究存在根本性差距——例如，目前还没有办法将工程级别的物理信息编码到神经网络中（没有专门针对它的 VAE/Transformer），模拟工程解决方案直到最近才变得可行（2024 年出现了 GPU 运行的模拟器，其速度比之前的任何模拟器都快 100-1000 倍），而且最重要的是，这是一个需要大量领域知识的机器学习任务。
几个月前，我爱上了这个问题，并且相信现在是解决这个问题的时候了。数据稀缺问题可以通过强化学习（RL）来解决——最近在 RL 方面取得了一些进展，使其在较小的训练数据上也能保持稳定（参见 SimbaV2/BROnet），工程级别的模拟可以通过 PINOs（物理信息神经网络算子——类似于物理信息神经网络，但速度快 10-100 倍，且更精确）来实现，3D 检测/分割/生成模型也变得近乎完美。而这正是我们真正需要的。
我希望组建一个 4-10 人的团队来解决这个问题。
硬件工程如此重要的原因是，如果我们能够可靠地进行硬件工程，我们就可以扩大生产规模，从而降低成本，并改善人类的所有物质需求——更多的能源生产、实物商品、汽车、住房——所有需要大规模制造才能运作的领域。
再说一次，我正在寻找一个团队来解决这个问题：
1. 我本人是具身智能研究员，主要研究 RL，并有一些机械工程背景。
2. 一到两位计算机视觉专家，
3. 高性能计算工程师，负责 RL 环境等，
4. 任何希望做出贡献的 AI 研究人员。
这里也存在一个可以探索的市场机会，如果你愿意，也可以考虑进去。开发一个原型需要几个月到一年的时间。我做过调查研究，尽管这基本上还是一个空白领域，我们需要共同努力，将所有输入整合起来。我还创建了一个可用于训练的 RL 环境（目前是闭源的）。
让我们为一项技术奠定基础/创造一个产品，让数百万人受益！
如果你想加入，请评论/给我发邮件到 maksymriabov2004@gmail.com。如果你至少在上述某个领域发表过论文，欢迎加入。
查看原文
Hello HN, I'm talking to ML researchers here,
There is a pretty challenging yet unexplored problem in ML yet - hardware engineering.<p>So far, everything goes against us solving this problem - pretrain data is basically inexistent (no abundance like in NLP/computer vision), there are fundamental gaps in research in the area - e.g. there is no way to encode engineering-level physics information into neural nets (no specialty VAEs/transformers oriented for it), simulating engineering solutions was very expensive up until recently (there are 2024 GPU-run simulators which run 100-1000x faster than anything before them), and on top of it it’s a domain-knowledge heavy ML task.<p>I’ve fell in love with the problem a few months ago, and I do believe that now is the time to solve this problem. The data scarcity problem is solvable via RL - there were recent advancements in RL that make it stable on smaller training data (see SimbaV2/BROnet), engineering-level simulation can be done via PINOs (Physics Informed Neural Operators - like physics-informed NNs, but 10-100x faster and more accurate), and 3d detection/segmentation/generation models are becoming nearly perfect. And that’s really all we need.
I am looking to gather a team of 4-10 people that would solve this problem.<p>The reason hardware engineering is so important is that if we reliably engineer hardware, we get to scale up our manufacturing, where it becomes much cheaper and we improve on all physical needs of the humanity - more energy generation, physical goods, automotive, housing - everything that uses mass manufacturing to work.<p>Again, I am looking for a team that would solve this problem:
1. I am an embodied AI researcher myself, mostly in RL and coming from some MechE background.
2. One or two computer vision people,
3. High-performance compute engineer for i.e. RL environments,
4. Any AI researchers who want to contribute.<p>There is also a market opportunity that can be explored too, so count that in if you wish. It will take a few months to a year to come up with a prototype. I did my research, although that’s basically an empty field yet, and we’ll need to work together to hack together all the inputs. I have also a created RL environment that could be used for training (currently closed-source).<p>Let us lay the foundation for a technology/create a product that would could benefit millions of people!<p>Comment if you want to join/mail me to maksymriabov2004@gmail.com. Everybody is welcome if you have at least published a paper in some of the aforementioned areas.