Search for: All records

The parameter space for any fixed architecture of feedforward ReLU neural networks serves as a proxy during training for the associated class of functions - but how faithful is this representation? It is known that many different parameter settings $\theta$ can determine the same function $f$. Moreover, the degree of this redundancy is inhomogeneous: for some networks, the only symmetries are permutation of neurons in a layer and positive scaling of parameters at a neuron, while other networks admit additional hidden symmetries. In this work, we prove that, for any network architecture where no layer is narrower than the input, there exist parameter settings with no hidden symmetries. We also describe a number of mechanisms through which hidden symmetries can arise, and empirically approximate the functional dimension of different network architectures at initialization. These experiments indicate that the probability that a network has no hidden symmetries decreases towards 0 as depth increases, while increasing towards 1 as width and input dimension increase. 

Training a neural network requires choosing a suitable learning rate, which involves a trade-off between speed and effectiveness of convergence. While there has been considerable theoretical and empirical analysis of how large the learning rate can be, most prior work focuses only on late-stage training. In this work, we introduce the maximal initial learning rate - the largest learning rate at which a randomly initialized neural network can successfully begin training and achieve (at least) a given threshold accuracy. Using a simple approach to estimate the maximal initial learning rate, we observe that in constant-width fully-connected ReLU networks, the maximal initial learning rate behaves differently from the maximum learning rate later in training. Specifically, we find that the maximal initial learning rate is well predicted as a power of depth times width, provided that (i) the width of the network is sufficiently large compared to the depth, and (ii) the input layer is trained at a relatively small learning rate. We further analyze the relationship between the maximal initial learning rate and the sharpness of the network at initialization, indicating they are closely though not inversely related. We formally prove bounds for the maximal initial learning rate in terms of depth times width that align with our empirical results. 

In late December 1973, the United States enacted what some would come to call “the pitbull of environmental laws.” In the 50 years since, the formidable regulatory teeth of the Endangered Species Act (ESA) have been credited with considerable successes, obliging agencies to draw upon the best available science to protect species and habitats. Yet human pressures continue to push the planet toward extinctions on a massive scale. With that prospect looming, and with scientific understanding ever changing,Scienceinvited experts to discuss how the ESA has evolved and what its future might hold.—Brad Wible

 

In late December 1973, the United States enacted what some would come to call “the pitbull of environmental laws.” In the 50 years since, the formidable regulatory teeth of the Endangered Species Act (ESA) have been credited with considerable successes, obliging agencies to draw upon the best available science to protect species and habitats. Yet human pressures continue to push the planet toward extinctions on a massive scale. With that prospect looming, and with scientific understanding ever changing, Science invited experts to discuss how the ESA has evolved and what its future might hold. 

In 1943, McCulloch and Pitts introduced a discrete recurrent neural network as a model for computation in brains. The work inspired breakthroughs such as the first computer design and the theory of finite automata. We focus on learning in Hopfield networks, a special case with symmetric weights and fixed-point attractor dynamics. Specifically, we explore minimum energy flow (MEF) as a scalable convex objective for determining network parameters. We catalog various properties of MEF, such as biological plausibility, and then compare to classical approaches in the theory of learning. Trained Hopfield networks can perform unsupervised clustering and define novel error-correcting coding schemes. They also efficiently find hidden structures (cliques) in graph theory. We extend this known connection from graphs to hypergraphs and discover n-node networks with robust storage of 2Ω(n1−ϵ) memories for any ϵ>0. In the case of graphs, we also determine a critical ratio of training samples at which networks generalize completely.