Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
This paper shows that GD/SGD can minimize the training loss of RNNs with linear convergence rate assuming the hidden layer width is sufficiently large (polynomial in data size and time horizon length). In order to prove this, the authors show that within a small region around the initialization, the norm square of the gradient can be lower bounded by the function value (Theorem 3). The authors further show that the loss function is somewhat smooth (Theorem 4), which guarantees that moving in the negative gradient direction can decrease the function value. This paper builds new techniques to analyze multi-layer ReLU networks. This paper shows that with appropriate initialization, ReLU activations avoid exponential exploding and exponential vanishing. This paper also shows that within a small region around the initialization, the multi-layer networks is pretty ``smooth’’. These techniques are very useful in the analysis of multi-layer networks and have been used in many following works. This paper is an important step towards the optimization theory of RNNs. The RNNs is much harder to analyze because the same recurrent unit is repeated applied and at initialization the randomness is shared across layers. This paper uses randomness decoupling techniques to analyze the spectral norm of RNNs at initialization. These techniques can be useful in other problems of RNNs. Overall this is a strong theory paper proving that GD/SGD can optimize RNNs in the over-parameterized setting. The techniques developed in this paper can be very useful in the analysis of multi-layer ReLU nets. My only concern is that the required hidden layer width is a polynomial in the number of samples (might be high order polynomial), which is not very practical. Also, the step size is very small, and the total movement of the weights is very small. In practice, the step size is much larger, and the weights move a lot. So, in some sense, this theory cannot explain the success of over-parameterization in practice. Reducing the dependency on m might require a very different idea. Here are some minor comments: 1. Line 142: shouldn’t it be h = D(Wh + Ax)? Ax is missing here. 2. It might be good to add a figure to illustrate the network architecture if space allows. -------------------------------------------- I have read the authors' response and other reviews. The authors have partially addressed my concerns on the network width and step size. I agree that as long as we assume the data is generated by some simple model, the requirement on the network width can be significantly reduced. Regarding the step size, the authors argue that this work can give some intuitions on the second phase of NN training when the step size decays and training loss goes to zero. However, the weights have moved a lot in the first phase (when the step size is large) and are not random anymore, it's not clear whether the current techniques can still work after phase 1. Despite these limitations, I still think this is a good theory paper and I will keep my score.
The paper proves that with overparameterized RNNs (polynomial in the training data), GD/SGD can find a global minima in polynomial time, with high probability, if the neural network is initialized according to some distribution (and is based on a three-layer RNN with ReLU activation functions). The paper does not eliminate the existence of worse local minima for proving these results, which means even if there are bad local minima, with high probability GD/SGD can still find a global minima. This could explain why in practice overparameterized networks perform better. Overall I feel the paper helps understanding how neural network works in practice. It can be better if an easy to understand intuition of why overparameterization helps (both proof and practical problems) is provided. ==== The authors have answered my concern on better explaining the intuition. I'll keep my score.
This work studies optimization of $\ell_2$-loss for training multilayer nonlinear recurrent neural network. The paper is well-written, well-presented, and easy to follow. When each layer of the neural network is highly over-parameterized, with several additional assumptions, it proves that the (stochastic) gradient descents converges linearly to zero training loss. The overall theoretical result is impressive, and several new results have been developed recently based on this work. The major concern is the practicality and generality of the result for real applications. Below are some more detailed comments: 1. However, the major concern is the practicality of the assumptions and hence the results. The theory seems to suggest that random initialization is already ``good’’ enough, that one only needs to make small adjustment of the weights to obtain zero training loss. Recent work [L] Lenaic Chizat, Edouard Oyallon, Francis Bach, ``On Lazy Training in Differentiable Programming’’, 2018. [G] Gilad Yehudai and Ohad Shamir. On the power and limitations of random features for understanding neural networks. arXiv preprint arXiv:1904.00687, 2019. explains that the regime (This is a regime where small changes of the weights result in large change of function values.) of the model the authors considered in this submission tends to behave like linear models (where they call lazy training), which seems to be not yet sufficient to explain the success of deep neural network in practice (demonstrated by experimental results on deep CNN). The authors should at least cite these results, and provide a detailed discussion on this. 2. The original idea of this line of work is coming from the neuron tangent kernel. This is from infinite width to finite width (asymptotic to non-asymptotic). The authors need to cite and recognize their work Arthur Jacot, Franck Gabriel, Clément Hongler. Neural Tangent Kernel: Convergence and Generalization in Neural Networks. 3. It seems quite strange to the reviewer that the matrix A and B can be set to their random initialization without optimization, and the training loss can be optimized to 0. Under the considered setting here, it seems that random initializations are close to the optimal solution. In the result, it seems that gradient descent on the weights W (given the stepsize eta very small in the theory) makes very little adjustment (the properties in Theorem 3 and Theorem 4 only hold when W is very close to the initialization), but producing zero training loss. 4. In the main theorem, the authors hide the degree of polynomial of the overparameterization in m. The degree of the polynomial seems to be very high. Can the author give us a sense how loose are the bounds, and what is the conjecture dependence for the result to hold here? 5. References. It would be great if the authors can make the citations in order of appearance and abbreviate them (e.g., page 1, change [58,19] to [1-2], and [7,49,54,31,15,18,36,60,59] should be abbreviated [3-11], etc). ====== After Rebuttal ======= The rebuttal clarifies and addresses most of my concerns, especially the relation of this work to [L], [G], and [J]. Albeit its practicality (it requires significant overparameterization), I agree with other reviewers that this is a solid theory paper that has triggered a lot interest in recent theoretical understandings of deep neural networks (e.g., SGD on the training loss of multilayer ReLU network). It also provides some explanation how SGD works in the early stages of training neural network.