转换不连续函数以用于神经网络
是否有任何转换不连续函数/值以用于神经网络之类的通用方法?
例如:
角度从 2pi -> 不连续0,尽管值本质上是相同的。在这种情况下,我能想到的最简单的变换是将角度转换为 3 个均匀间隔的角度(0、2/3pi、4/3pi)的一组最小角距离。
在这种情况下,两个不同的角度不应产生相同的三个距离,但两个非常相似的角度应始终产生三个非常相似的距离。
虽然我还没有对此进行广泛的测试,但当尝试使用最小复杂度的分类器(例如单层神经网络)时,该变换似乎更合适。
我想知道这种转换风格是否有任何一种“一般形式”可以应用于其他情况。
Are there any general methods of transforming discontinuous functions/values for use in something like a neural network?
For example:
Angles are discontinuous from 2pi -> 0, even though the values are essentially the same. The simplest transform I can think of in this case is to convert the angle into a set of minimum angular distances from 3 evenly spaced angles (0, 2/3pi, 4/3pi).
In that case, no two different angles should result in the same three distances, yet two very similar angles should always result in three very similar distances.
While I haven't extensively tested this, the transform seems more suitable when trying to use a minimally complex classifier (such as a single layer neural network).
I'm wondering if there is any sort of 'general form' of this style of transformation that could be applied to other situations.
如果你对这篇内容有疑问,欢迎到本站社区发帖提问 参与讨论,获取更多帮助,或者扫码二维码加入 Web 技术交流群。
绑定邮箱获取回复消息
由于您还没有绑定你的真实邮箱,如果其他用户或者作者回复了您的评论,将不能在第一时间通知您!
发布评论
评论(1)
为什么不使用 sin(x) 和 cos(x)? (这只是我想到的第一个想法)
Why not using sin(x) and cos(x)? (It's just the first idea that came to my mind)