Deep Learners Benefit More from Out-of-Distribution Examples

نویسندگان

  • Yoshua Bengio
  • Frédéric Bastien
  • Arnaud Bergeron
  • Nicolas Boulanger-Lewandowski
  • Thomas M. Breuel
  • Youssouf Chherawala
  • Moustapha Cissé
  • Myriam Côté
  • Dumitru Erhan
  • Jeremy Eustache
  • Xavier Glorot
  • Xavier Muller
  • Sylvain Pannetier Lebeuf
  • Razvan Pascanu
  • Salah Rifai
  • François Savard
  • Guillaume Sicard
چکیده

Recent theoretical and empirical work in statistical machine learning has demonstrated the potential of learning algorithms for deep architectures, i.e., function classes obtained by composing multiple levels of representation. The hypothesis evaluated here is that intermediate levels of representation, because they can be shared across tasks and examples from different but related distributions, can yield even more benefits. Comparative experiments were performed on a large-scale handwritten character recognition setting with 62 classes (upper case, lower case, digits), using both a multi-task setting and perturbed examples in order to obtain out-ofdistribution examples. The results agree with the hypothesis, and show that a deep learner did beat previously published results and reached human-level performance.

برای دانلود رایگان 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

Deep Self-Taught Learning for Handwritten Character Recognition

Recent theoretical and empirical work in statistical machine learning has demonstrated the importance of learning algorithms for deep architectures, i.e., function classes obtained by composing multiple non-linear transformations. Self-taught learning (exploiting unlabeled examples or examples from other distributions) has already been applied to deep learners, but mostly to show the advantage ...

متن کامل

Complimenting Functions by Native English Speakers and Iranian EFL Learners: A Divergence or Convergence

The study of compliment speech act has been under investigation on many occasions in recent years. In this study, an attempt is made to explore appraisals performed by native English speakers and Iranian EFL learners to find out how these two groups diverge or converge from each other with regard to complimenting patterns and norms. The participants of the study were 60 advanced Iranian EFL lea...

متن کامل

The Effect of L1 Persian on the Acquisition of English L2 Orthographic System on the Shared Grounds

This paper elaborates on Persian and English orthographic shared aspects to study the effects of L1 Persian on learning English as a foreign language. While there are some examples of letter and sound mismatches in the orthographic system of both languages, those of English are more complex than Persian. In order to see the effect of the mismatch between orthography and transcription, 40 Persia...

متن کامل

Implicit and Explicit Instruction and EFL Learners’ Implicit Knowledge Development: Evidence from Word Monitoring Task

Research on the effect of implicit and explicit instruction on developing learners’ explicit knowledge by the use of measures of explicit knowledge abounds in the literature. However, measuring learners’ implicit knowledge employing fine-grained measures has rarely been the concern of researchers in the field. Therefore, the present study is an attempt to scrutinize the effectiveness of implici...

متن کامل

A novel method based on a combination of deep learning algorithm and fuzzy intelligent functions in order to classification of power quality disturbances in power systems

Automatic classification of power quality disturbances is the foundation to deal with power quality problem. From the traditional point of view, the identification process of power quality disturbances should be divided into three independent stages: signal analysis, feature selection and classification. However, there are some inherent defects in signal analysis and the procedure of manual fe...

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

عنوان ژورنال:

دوره   شماره 

صفحات  -

تاریخ انتشار 2011