University of Waterloo

About: University of Waterloo is a education organization based out in Waterloo, Ontario, Canada. It is known for research contribution in the topics: Population & Context (language use). The organization has 36093 authors who have published 93906 publications receiving 2948139 citations. The organization is also known as: UW & uwaterloo.

On Calibration of Modern Neural Networks

Abstract: Confidence calibration -- the problem of predicting probability estimates representative of the true correctness likelihood -- is important for classification models in many applications. We discover that modern neural networks, unlike those from a decade ago, are poorly calibrated. Through extensive experiments, we observe that depth, width, weight decay, and Batch Normalization are important factors influencing calibration. We evaluate the performance of various post-processing calibration methods on state-of-the-art architectures with image and document classification datasets. Our analysis and experiments not only offer insights into neural network learning, but also provide a simple and straightforward recipe for practical settings: on most datasets, temperature scaling -- a single-parameter variant of Platt Scaling -- is surprisingly effective at calibrating predictions.

On calibration of modern neural networks

Abstract: Confidence calibration -- the problem of predicting probability estimates representative of the true correctness likelihood -- is important for classification models in many applications. We discover that modern neural networks, unlike those from a decade ago, are poorly calibrated. Through extensive experiments, we observe that depth, width, weight decay, and Batch Normalization are important factors influencing calibration. We evaluate the performance of various post-processing calibration methods on state-of-the-art architectures with image and document classification datasets. Our analysis and experiments not only offer insights into neural network learning, but also provide a simple and straightforward recipe for practical settings: on most datasets, temperature scaling -- a single-parameter variant of Platt Scaling -- is surprisingly effective at calibrating predictions.

Establishing a causal chain: why experiments are often more effective than mediational analyses in examining psychological processes.

Abstract: The authors propose that experiments that utilize mediational analyses as suggested by R. M. Baron and D. A. Kenny (1986) are overused and sometimes improperly held up as necessary for a good social psychological paper. The authors argue that when it is easy to manipulate and measure a proposed psychological process that a series of experiments that demonstrates the proposed causal chain is superior. They further argue that when it is easy to manipulate a proposed psychological process but difficult to measure it that designs that examine underlying process by utilizing moderation can be effective. It is only when measurement of a proposed psychological process is easy and manipulation of it is difficult that designs that rely on mediational analyses should be preferred, and even in these situations careful consideration should be given to the limiting factors of such designs.

Weyl semimetal in a topological insulator multilayer.

Abstract: We propose a simple realization of the three-dimensional (3D) Weyl semimetal phase, utilizing a multilayer structure, composed of identical thin films of a magnetically doped 3D topological insulator, separated by ordinary-insulator spacer layers. We show that the phase diagram of this system contains a Weyl semimetal phase of the simplest possible kind, with only two Dirac nodes of opposite chirality, separated in momentum space, in its band structure. This Weyl semimetal has a finite anomalous Hall conductivity and chiral edge states and occurs as an intermediate phase between an ordinary insulator and a 3D quantum anomalous Hall insulator. We find that the Weyl semimetal has a nonzero dc conductivity at zero temperature, but Drude weight vanishing as ${T}^{2}$, and is thus an unusual metallic phase, characterized by a finite anomalous Hall conductivity and topologically protected edge states.