Translated Learning Wenyuan Dai Yuqiang Chen GuiRong Xue

Translated Learning Wenyuan Dai, Yuqiang Chen, Gui-Rong Xue, Qiang Yang, and Yong Yu. Translated Learning. In Proceedings of Twenty. Second Annual Conference on Neural Information Processing Systems (NIPS 2008), December 8, 2008, Vancouver, British Columbia, Canada.

Definition • Translated Learning – Learning across Different Feature Spaces

Applications • Cross-language Classification – Rigutini et al. , WI 2005; Ling et al. , WWW 2008; … • Text-aided Image Classification – We submitted two papers to AAAI 2008 & ICML 2008 respectively. • Future work – Text to Music – Text to Video – Image to Video –…

Related Work • Cross-language Classification – Rigutini et al. , WI 2005 • English to Italian – Ling et al. , WWW 2008 • English to Chinese • Most cross-language classification approaches relies on machine translation. – Ad hoc – Machine translation is difficult in most scenarios. • E. g. text-to-picture translation

Basic Idea • Instance-level machine translation relies on understanding instances, at least basically. – Machine translation in NLP is an easy special case, since it is based on sentence understanding. • Classification models are usually on the featurelevel. • Translating classification models is also on the feature-level. – could be much easier than instance-level translation

Human Learning Example • Task: tyrannosaurus vs stegosaurus – htyrannosaurus: bipedal carnivore with a massive skull balanced by long, heavy tail. Its forelimbs were small and retained only two digits. – stegosaurus: quadruped ornithischian dinosaur of four long bony spikes on a flexible tail and two rows of upright triangular bony plates running along the back.

Model-level Translation Elephants are big mammals on earth. . . Input massive hoofed mammal of Africa. . . Learning Output translating learning models Input Learning Output

Naive Bayesian Approach difficult to estimate • Incorporating translator

Risk Minimization Approach • Loss function

Inference • Assume there is no prior difference among all the classes

Model Estimation • KL-divergence • Negative of cosine • Negative of PCC

Experimental Results

Experimental Results

Outline • • Introduction Related Work Our Research Future Work

Future Work • More applications – Cross-language classification • Using dictionaries as translators – Text to music, video, … – Image to video • Improving translator estimation – Integrating text classification and translator estimation into one optimization model

Questions