4 ms·
What I knew was that local minima are not that problematic when the state space is highly dimensional. Only saddle points appear in high dimensional spaces. The
by visarga 10y ago
What I knew was that local minima are not that problematic when the state space is highly dimensional. Only saddle points appear in high dimensional spaces. The 2D example is not realistic.
Here is a quote from a paper I randomly sampled on arxiv:
> For an expected loss function of a deep nonlinear neural network, we prove the following statements under the independence assumption adopted from recent work: 1) the function is non-convex and non-concave, 2) every local minimum is a global minimum, 3) every critical point that is not a global minimum is a saddle point, and 4) the property of saddle points differs for shallow networks (with three layers) and deeper networks (with more than three layers).
https://arxiv.org/abs/1605.07110v1 https://arxiv.org/abs/1605.07110v1
Also, if your critique is related to the perceived lacks of backpropagation, keep in mind than reinforcement learning is also a kind of backpropagation of a reward, but this time the reward is much sparser and low dimensional. Thus, they are somewhere in-between supervised and unsupervised learning, not quite enjoying the full supervision of backpropagating at every example, but still learning based on an external critic.
The way forward is to implement reinforcement learning agents with memory and attention. These systems are neural turing machines, they can compute in a sequence of steps.
- kennethfriedman 10y agoAuthor here, thanks for the comment! While this is an interesting read (and NTMs are great), it is not particularly relevant to the model I describe or this paper. I think you are misunderstanding my generality of my paper: I am not discussing a particular method of deep learning. I am using the idea of gradient descent as an metaphor for the field of AI, itself. As described in the second paragraph of the "Gradient Descent" section, this analogy is not high dimensional. In fact, it is only three dimensional: distance the field is from General AI, time, and a hypothetical "method of attack".
- phreeza 10y agoI think the reasoning still holds. If the dimensions are reasonably independent of each other, the probability of a local minimum occurring in any kind of high-dimensional surface falls rapidly as the number of dimensions increases. And I would argue that machine learning methods are a pretty high-dimensional search space, describing them in a single 'method of attack' dimension is probably not right.
- jal278 10y agoBut by that line of reasoning, shouldn't we never hit dead ends in AI research at all -- why has AI progress been so difficult, then? Wouldn't any field of research with many dimensions of variation never get stuck on its path towards its ultimate goals, ever? Couldn't different objective functions be structurally more difficult than others to optimize? No matter how high-dimensional the search-space, trying to create a gaming laptop in the middle ages would have been a pretty frustrating experience.
- phreeza 10y agoBecause there are plenty of saddle points. Gradient descent slows down quite a bit at those, too.
- dharma1 10y ago"reinforcement learning agents with memory and attention" Here's a hybrid approach https://arxiv.org/abs/1605.07427 https://arxiv.org/abs/1605.07427