让语言模型「轻装上阵」。
第一种策略是使用初等运算代替 MatMul,例如,在卷积神经网络 (CNN) 中,用有符号加法代替乘法; 第二种方法是使用二值或三值化量化,将 MatMul 值在累加之前要么翻转要么清零。比如脉冲神经网络 (SNN) 使用二值激活,而二值化网络 BNN 使用量化权重。
![图片](https://image.jiqizhixin.com/uploads/editor/4e0cbcf8-b2fd-4bc2-9eed-fb9dd8026711/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/281a62df-675d-465d-8458-69e5913b60e5/640.png)
论文地址:https://arxiv.org/pdf/2406.02528 项目地址:https://github.com/ridgerchu/matmulfreellm 论文标题:Scalable MatMul-free Language Modeling
![图片](https://image.jiqizhixin.com/uploads/editor/078cd91f-48ff-40e3-a09d-ab01eba7bf6d/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/67ca68b4-f7e6-4ae1-8486-05cf560aa73e/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/6cc7d6f8-dfc3-444e-bf99-fd453f64bdf4/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/a2f3cd22-470c-4387-abdf-b850023620b4/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/2948f91b-43b2-4644-89d8-36963599a673/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/05d4d2f6-cb6d-44fc-9abb-3a51c2038273/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/1f2e4dec-603e-4ecf-9ec4-dc3ae4fbc5bb/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/dcd20c99-18a2-4921-a16d-e70f643e8b03/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/974136ba-de95-4e6d-b11e-fef757545122/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/dfaf90a6-0222-48e0-980b-e93ada6c0e4b/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/a32551c9-b450-44cf-b3a4-8bef1ee2e358/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/50ef0526-2f32-48e8-b298-375f73d5301c/640.png)