DiverseNet: When One Right Answer is not Enough

نویسندگان

  • Michael Firman
  • Neill D. F. Campbell
  • Lourdes Agapito
  • Gabriel J. Brostow
چکیده

Many structured prediction tasks in machine vision have a collection of acceptable answers, instead of one definitive ground truth answer. Segmentation of images, for example, is subject to human labeling bias. Similarly, there are multiple possible pixel values that could plausibly complete occluded image regions. State-of-the art supervised learning methods are typically optimized to make a single test-time prediction for each query, failing to find other modes in the output space. Existing methods that allow for sampling often sacrifice speed or accuracy. We introduce a simple method for training a neural network, which enables diverse structured predictions to be made for each test-time query. For a single input, we learn to predict a range of possible answers. We compare favorably to methods that seek diversity through an ensemble of networks. Such stochastic multiple choice learning faces mode collapse, where one or more ensemble members fail to receive any training signal. Our best performing solution can be deployed for various tasks, and just involves small modifications to the existing single-mode architecture, loss function, and training regime. We demonstrate that our method results in quantitative improvements across three challenging tasks: 2D image completion, 3D volume estimation, and flow prediction.

برای دانلود متن کامل این مقاله و بیش از 32 میلیون مقاله دیگر ابتدا ثبت نام کنید

ثبت نام

اگر عضو سایت هستید لطفا وارد حساب کاربری خود شوید

منابع مشابه

A Study of Answer Changing Behavior in MC Tests: The Effect of the Academic Course, Field of Study, Gender and Teachers’ Attitude

This study investigated the answer changing behavior of the Iranian university students and its relationship to the academic course, gender, and field of study. 362 students at the University of Isfahan took part in the study. 76 English language teachers were also surveyed for their attitude toward answer changing on MC tests. The results indicated that 42.11% of the teachers had a negative at...

متن کامل

I-43: Scientific and Religious Controversies on The Beginning of Human Life- What Does 3D/4D Sonography Offer?

One of the most controversial topics in modern bioethics, science, and philosophy is the beginning of individual human life. In the seemingly endless debate, strongly stimulated by recent technologic advances in human reproduction, a synthesis between scientific data and hypothesis, philosophical thought, and issues of humanities has become a necessity to deal with ethical, juridical, and socia...

متن کامل

ارائه روشی پویا جهت پاسخ به پرس‌وجوهای پیوسته تجمّعی اقتضایی

Data Streams are infinite, fast, time-stamp data elements which are received explosively. Generally, these elements need to be processed in an online, real-time way. So, algorithms to process data streams and answer queries on these streams are mostly one-pass. The execution of such algorithms has some challenges such as memory limitation, scheduling, and accuracy of answers. They will be more ...

متن کامل

Joseph Kovac

The finite difference method represents a highly straightforward and logical approach to approximating continuous problems using discrete methods. At its heart is a simple idea: substitute finite, discrete differences for derivatives in some way appropriate for a given problem, make the time and space steps the right size, run the difference method, and get an approximation of the answer. Many ...

متن کامل

P-53: Civil Responsibility of Parties to a Partial Surrogacy Contract

Background: During recent years, scientists have presented several methods for treating infertility; these methods occasionally are highly efficient for infertile couples but can be concerning because of legal and ethical issues. One of these methods is surrogacy. Surrogacy is used when the main etiology of infertility is in the uterus caused by these three reasons: lack of uterine‚ congenital ...

متن کامل

ذخیره در منابع من


  با ذخیره ی این منبع در منابع من، دسترسی به آن را برای استفاده های بعدی آسان تر کنید

برای دانلود متن کامل این مقاله و بیش از 32 میلیون مقاله دیگر ابتدا ثبت نام کنید

ثبت نام

اگر عضو سایت هستید لطفا وارد حساب کاربری خود شوید

عنوان ژورنال:

دوره   شماره 

صفحات  -

تاریخ انتشار 2018