In this paper, we prove a conjecture published in 1989 and also partially address an open problem announced at the Conference on Learning Theory (COLT) 2015. 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). Moreover, we prove that the same four statements hold for deep linear neural networks with any depth, any widths and no unrealistic assumptions. As a result, we present an instance, for which we can answer to the following question: how difficult to directly train a deep model in theory? It is more difficult than the classical machine learning models (because of the non-convexity), but not too difficult (because of the nonexistence of poor local minima and the property of the saddle points). We note that even though we have advanced the theoretical foundations of deep learning, there is still a gap between theory and practice.

--Inherent in virtually every iterative machine learning algorithm is the problem of hyper-parameter tuning which includes three major design parameters: (a) the complexity of the model, e.g., the number of neurons in a neural network, (b) the initial conditions, which heavily affect the behavior of the algorithm, and (c) the dissimilarity measure used to quantify its performance. We introduce an online prototype-based learning algorithm that can be viewed as a progressively growing competitive-learning neural network architecture for classification and clustering. The learning rule of the proposed approach is formulated as an online gradient-free stochastic approximation algorithm that solves a sequence of appropriately defined optimization problems, simulating an annealing process. The annealing nature of the algorithm contributes to avoiding poor local minima, offers robustness with respect to the initial conditions, and provides a means to progressively increase the complexity of the learning model, through an intuitive bifurcation phenomenon. The proposed approach is interpretable, requires minimal hyper-parameter tuning, and allows online control over the performance-complexity trade-off. Finally, we show that Bregman divergences appear naturally as a family of dissimilarity measures that play a central role in both the performance and the computational complexity of the learning algorithm. EARNING from data samples has become an important component of artificial intelligence. While virtually all learning problems can be formulated as constrained stochastic optimization problems, the optimization methods can be intractable, typically dealing with mixed constraints and very large, or even infinite-dimensional spaces [1]. For this reason, feature extraction, model selection and design, and analysis of optimization methods, have been the cornerstone of machine learning algorithms from their genesis until today. Deep learning methods, currently dominating the field of machine learning due to their performance in multiple applications, attempt to learn feature representations from data, using biologically-inspired models in artificial neural networks [2], [3]. Manuscript published in the IEEE Transactions on Neural Networks and Learning Systems (TNNLS).

The paper makes a significant dent in the central mystery of deep learning: why are neural networks so easy to optimize? The precision of the linear result is quite impressive! The nonlinear result still makes a implausible assumption, but is stronger than previous work and the implausibility is carefully discussed. The paper itself was a delight to read, though I will admit that the full proofs were a bit of a slog. My main question is whether the proofs could be simplified.

We study the benefits of complex-valued weights for neural networks. We prove that shallow complex neural networks with quadratic activations have no spurious local minima. In contrast, shallow real neural networks with quadratic activations have infinitely many spurious local minima under the same conditions. In addition, we provide specific examples to demonstrate that complex-valued weights turn poor local minima into saddle points. The activation function CReLU is also discussed to illustrate the superiority of analytic activations in complex-valued neural networks.

In this paper, we prove a conjecture published in 1989 and also partially address an open problem announced at the Conference on Learning Theory (COLT) 2015. 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). Moreover, we prove that the same four statements hold for deep linear neural networks with any depth, any widths and no unrealistic assumptions. As a result, we present an instance, for which we can answer to the following question: how difficult to directly train a deep model in theory? It is more difficult than the classical machine learning models (because of the non-convexity), but not too difficult (because of the nonexistence of poor local minima and the property of the saddle points). We note that even though we have advanced the theoretical foundations of deep learning, there is still a gap between theory and practice.

Over the past decade, advances in deep learning have transformed the fortunes of the artificial intelligence (AI) community. The neural network approach that researchers had largely written off by the end of the 1990s now seems likely to become the most widespread technology in machine learning. However, protagonists find it difficult to explain why deep learning often works well, but is prone to seemingly bizarre failures. The success of deep learning came with rapid improvements in computational power that came through the development of highly parallelized microprocessors and the discovery of ways to train networks with enormous numbers of virtual neurons assembled into tens of linked layers. Before these advances, neural networks were limited to simple structures that were easily outclassed in image and audio classification tasks by other machine-learning architectures such as support vector machines.

The loss surface of deep neural networks has recently attracted interest in the optimization and machine learning communities as a prime example of high-dimensional non-convex problem. Some insights were recently gained using spin glass models and mean-field approximations, but at the expense of strongly simplifying the nonlinear nature of the model. In this work, we do not make any such assumption and study conditions on the data distribution and model architecture that prevent the existence of bad local minima. Our theoretical work quantifies and formalizes two important \emph{folklore} facts: (i) the landscape of deep linear networks has a radically different topology from that of deep half-rectified ones, and (ii) that the energy landscape in the non-linear case is fundamentally controlled by the interplay between the smoothness of the data distribution and model over-parametrization. Our main theoretical contribution is to prove that half-rectified single layer networks are asymptotically connected, and we provide explicit bounds that reveal the aforementioned interplay. The conditioning of gradient descent is the next challenge we address. We study this question through the geometry of the level sets, and we introduce an algorithm to efficiently estimate the regularity of such sets on large-scale networks. Our empirical results show that these level sets remain connected throughout all the learning phase, suggesting a near convex behavior, but they become exponentially more curvy as the energy level decays, in accordance to what is observed in practice with very low curvature attractors.