Initialization of deep networks
As we all know, the solution to a non-convex optimization algorithm (like stochastic gradient descent) depends on the initial values of the parameters. This post is about choosing initialization parameters for deep networks and how it affects the convergence. We will also discuss the related topic of vanishing gradients.
First, let's go back to the time of sigmoidal activation functions and initialization of parameters using IID Gaussian or uniform distributions with fairly arbitrarily set variances. Building deep networks was difficult because of exploding or vanishing activations and gradients. Let's take activations first: If all your parameters are too small, the variance of your activations will drop in each layer. This is a problem if your activation function is sigmoidal, since it is approximately linear close to 0. That is, you gradually lose your non-linearity, which means there is no benefit to having multiple layers. If, on the other hand, your activations become larger and larger, then your activations will saturate and become meaningless, with gradients approaching 0.
Let us consider one layer and forget about the bias. Note that the following analysis and conclussion is taken from Glorot and Bengio[1]. Consider a weight matrix W∈Rm×n, where each element was drawn from an IID Guassian with variance Var(W). Note that we are a bit abusive with notation letting W denote both a matrix and a univariate random variable. We also assume there is no correlation between our input and our weights and both are zero-mean. If we consider one filter (row) in W, say w (a random vector), then the variance of the output signal over the input signal is:
As we build a deep network, we want the variance of the signal going forward in the network to remain the same, thus it would be advantageous if nVar(W)=1. The same argument can be made for the gradients, the signal going backward in the network, and the conclusion is that we would also like mVar(W)=1. Unless n=m, it is impossible to sastify both of these conditions. In practice, it works well if both are approximately satisfied. One thing that has never been clear to me is why it is only necessary to satisfy these conditions when picking the initialization values of W. It would seem that we have no guarantee that the conditions will remain true as the network is trained.
Nevertheless, this Xavier initialization (after Glorot's first name) is a neat trick that works well in practice. However, along came rectified linear units (ReLU), a non-linearity that is scale-invariant around 0 and does not saturate at large input values. This seemingly solved both of the problems the sigmoid function had; or were they just alleviated? I am unsure of how widely used Xavier initialization is, but if it is not, perhaps it is because ReLU seemingly eliminated this problem.
However, take the most competative network as of recently, VGG[2]. They do not use this kind of initialization, although they report that it was tricky to get their networks to converge. They say that they first trained their most shallow architecture and then used that to help initialize the second one, and so forth. They presented 6 networks, so it seems like an awfully complicated training process to get to the deepest one.
A recent paper by He et al.[3] presents a pretty straightforward generalization of ReLU and Leaky ReLU. What is more interesting is their emphasis on the benefits of Xavier initialization even for ReLU. They re-did the derivations for ReLUs and discovered that the conditions were the same up to a factor 2. The difficulty Simonyan and Zisserman had training VGG is apparently avoidable, simply by using Xavier intialization (or better yet the ReLU adjusted version). Using this technique, He et al. reportedly trained a whopping 30-layer deep network to convergence in one go.
Another recent paper tackling the signal scaling problem is by Ioffe and Szegedy[4]. They call the change in scale internal covariate shift and claim this forces learning rates to be unnecessarily small. They suggest that if all layers have the same scale and remain so throughout training, a much higher learning rate becomes practically viable. You cannot just standardize the signals, since you would lose expressive power (the bias disappears and in the case of sigmoids we would be constrained to the linear regime). They solve this by re-introducing two parameters per layer, scaling and bias, added again after standardization. The training reportedly becomes about 6 times faster and they present state-of-the-art results on ImageNet. However, I'm not certain this is the solution that will stick.
I reckon we will see a lot more work on this frontier in the next few years. Especially since it also relates to the -- right now wildly popular -- Recurrent Neural Network (RNN), which connects output signals back as inputs. The way you train such network is that you unroll the time axis, treating the result as an extremely deep feedforward network. This greatly exacerbates the vanishing gradient problem. A popular solution, called Long Short-Term Memory (LSTM), is to introduce memory cells, which are a type of teleport that allows a signal to jump ahead many time steps. This means that the gradient is retained for all those time steps and can be propagated back to a much earlier time without vanishing.
This area is far from solved, and until then I think I will be sticking to Xavier initialization. If you are using Caffe, the one take-away of this post is to use the following on all your layers:
weight_filler {
type: "xavier"
}
References
X. Glorot and Y. Bengio, “Understanding the difficulty of training deep feedforward neural networks,” in International conference on artificial intelligence and statistics, 2010, pp. 249–256.
K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014. [pdf]
K. He, X. Zhang, S. Ren, and J. Sun, “Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification,” arXiv:1502.01852 [cs], Feb. 2015. [pdf]
S. Ioffe and C. Szegedy, “Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift,” arXiv:1502.03167 [cs], Feb. 2015. [pdf]