Recent approaches to efficiently ensemble neural networks have shown that strong robustness and uncertainty performance can be achieved with a negligible gain in parameters over the original network. However, these methods still require multiple forward passes for prediction, leading to a significant computational cost. In this work, we show a surprising result: the benefits of using multiple predictions can be achieved `for free' under a single model's forward pass. In particular, we show that, using a multi-input multi-output (MIMO) configuration, one can utilize a single model's capacity to train multiple subnetworks that independently learn the task at hand. By ensembling the predictions made by the subnetworks, we improve model robustness without increasing compute. We observe a significant improvement in negative log-likelihood, accuracy, and calibration error on CIFAR10, CIFAR100, ImageNet, and their out-of-distribution variants compared to previous methods.

Emergent properties of the local geometry of neural loss landscapesStanislav Fort Surya Ganguli Stanford University Stanford, CA, USA Stanford University Stanford, CA, USA Abstract The local geometry of high dimensional neural network loss landscapes can both challenge our cherished theoretical intuitions as well as dramatically impact the practical success of neural network training. Indeed recent works have observed 4 striking local properties of neural loss landscapes on classification tasks: (1) the landscape exhibits exactly C directions of high positive curvature, where C is the number of classes; (2) gradient directions are largely confined to this extremely low dimensional subspace of positive Hessian curvature, leaving the vast majority of directions in weight space unexplored; (3) gradient descent transiently explores intermediate regions of higher positive curvature before eventually finding flatter minima; (4) training can be successful even when confined to low dimensional random affine hy-perplanes, as long as these hyperplanes intersect a Goldilocks zone of higher than average curvature. We develop a simple theoretical model of gradients and Hessians, justified by numerical experiments on architectures and datasets used in practice, that simultaneously accounts for all 4 of these surprising and seemingly unrelated properties. Our unified model provides conceptual insights into the emergence of these properties and makes connections with diverse topics in neural networks, random matrix theory, and spin glasses, including the neural tangent kernel, BBP phase transitions, and Derrida's random energy model. 1 Introduction The geometry of neural network loss landscapes and the implications of this geometry for both optimization and generalization have been subjects of intense interest in many works, ranging from studies on the lack of local minima at significantly higher loss than that of the global minimum [1, 2] to studies debating relations between the curvature of local minima and their generalization properties [3, 4, 5, 6]. Fundamentally, the neural network loss landscape is a scalar loss function over a very high D dimensional parameter space that could depend a priori in highly nontrivial ways on the very structure of real-world data itself as well as intricate properties of the neural network architecture. Moreover, the regions of this loss landscape explored by gradient descent could themselves have highly atypical geometric properties relative to randomly chosen points in the landscape.

There are many surprising and perhaps counter-intuitive properties of optimization of deep neural networks. We propose and experimentally verify a unified phenomenological model of the loss landscape that incorporates many of them. High dimensionality plays a key role in our model. Our core idea is to model the loss landscape as a set of high dimensional \emph{wedges} that together form a large-scale, inter-connected structure and towards which optimization is drawn. We first show that hyperparameter choices such as learning rate, network width and $L_2$ regularization, affect the path optimizer takes through the landscape in a similar ways, influencing the large scale curvature of the regions the optimizer explores. Finally, we predict and demonstrate new counter-intuitive properties of the loss-landscape. We show an existence of low loss subspaces connecting a set (not only a pair) of solutions, and verify it experimentally. Finally, we analyze recently popular ensembling techniques for deep networks in the light of our model.

We investigate neural network training and generalization using the concept of stiffness. We measure how stiff a network is by looking at how a small gradient step on one example affects the loss on another example. In particular, we study how stiffness varies with 1) class membership, 2) distance between data points (in the input space as well as in latent spaces), 3) training iteration, and 4) learning rate. We empirically study the evolution of stiffness on MNIST, FASHION MNIST, CIFAR-10 and CIFAR-100 using fully-connected and convolutional neural networks. Our results demonstrate that stiffness is a useful concept for diagnosing and characterizing generalization. We observe that small learning rates lead to initial learning of more specific features that do not translate well to improvements on inputs from all classes, whereas high learning rates initially benefit all classes at once. We measure stiffness as a function of distance between data points and observe that higher learning rates induce positive correlation between changes in loss further apart, pointing towards a regularization effect of learning rate. When training on CIFAR-100, the stiffness matrix exhibits a coarse-grained behavior suggestive of the model's awareness of super-class membership.

We explore the loss landscape of fully-connected neural networks using random, low-dimensional hyperplanes and hyperspheres. Evaluating the Hessian, $H$, of the loss function on these hypersurfaces, we observe 1) an unusual excess of the number of positive eigenvalues of $H$, and 2) a large value of $\mathrm{Tr}(H) / |H|$ at a well defined range of configuration space radii, corresponding to a thick, hollow, spherical shell we refer to as the \textit{Goldilocks zone}. We observe this effect for fully-connected neural networks over a range of network widths and depths on MNIST and CIFAR-10 with the $\mathrm{ReLU}$ non-linearity. The effect is not observed for the $\tanh$ non-linearity. Using our observations, we demonstrate a close connection between the Goldilocks zone, measures of local convexity/prevalence of positive curvature, and the suitability of a network initialization. We show that the high and stable accuracy reached when optimizing on random, low-dimensional hypersurfaces is directly related to the overlap between the hypersurface and the Goldilocks zone. We note that common initialization techniques initialize neural networks in this particular region of unusually high convexity, and offer a geometric intuition for their success. We take steps towards an analytic description of the general features of the loss function geometry, exploring its anisotropy and strong radial dependence. We support our theoretical results with experiments. Furthermore, we demonstrate that initializing a neural network at a number of points and selecting for high measures of local convexity such as $\mathrm{Tr}(H) / |H|$, number of positive eigenvalues of $H$, or low initial loss, leads to statistically significantly faster training on MNIST. Based on our observations, we hypothesize that the Goldilocks zone contains a high density of suitable initialization configurations.

We propose a novel architecture for $k$-shot classification on the Omniglot dataset. Building on prototypical networks, we extend their architecture to what we call Gaussian prototypical networks. Prototypical networks learn a map between images and embedding vectors, and use their clustering for classification. In our model, a part of the encoder output is interpreted as a confidence region estimate about the embedding point, and expressed as a Gaussian covariance matrix. Our network then constructs a direction and class dependent distance metric on the embedding space, using uncertainties of individual data points as weights. We show that Gaussian prototypical networks are a preferred architecture over vanilla prototypical networks with an equivalent number of parameters. We report state-of-the-art performance in 1-shot and 5-shot classification both in 5-way and 20-way regime (for 5-shot 5-way, we are comparable to previous state-of-the-art) on the Omniglot dataset. We explore artificially down-sampling a fraction of images in the training set, which improves our performance even further. We therefore hypothesize that Gaussian prototypical networks might perform better in less homogeneous, noisier datasets, which are commonplace in real world applications.