Embedded Files
Recommendation System Optimization
Recommendation System Optimization
There are several challenging tasks in content recommendation scenarios, including cold-start trial contents selection, personalized interest matching, etc. Performance optimization is even more critical when it comes to the professional news feed scene (e.g. a financial news feed APP).
Our optimizing solution includes: 1) analyzing contextual embedding features of historical/latest feed contents' similarity scores to determine trial items priority; 2) cold-start matching performance-boosting using online reinforcement learning models to efficiently leverage low-shot real-time interaction; 3) multi-stage flow resource allocation to efficiently deliver the high-quality contents 4) domain-expertized content/user tag structure for personalized interest matching.
NLP-based Text Understanding
NLP-based Text Understanding
Extractive text summarization: Open-source training data could potentially be insufficient for specialized domains like fund research reports in different industry domains. We propose a data-efficient solution by pretraining the major part of the model upon dataset across different domains and adding multi-head output layers for domain fine-tuning adaption.
Text style tuning: To enhance the expression style diversity and attraction, and to keep the rigorousness of text used in certain contexts, we select sentences from various real-life scenes to build up iterable corpora based on explicable metrics (e.g. how likely a title triggers a user click while browsing), which could be used as goal templates for current text modification generation. Meanwhile, a language model is involved as the scorer to determine the final modification.
Transfer Learning with Interactive Reinforcement Learning
Transfer Learning with Interactive Reinforcement Learning
Many learning methods, such as reinforcement learning, suffers from a slow beginning, especially in complicated domains. The motivation of transfer learning is to use limited prior knowledge to help learning agents bootstrap at the start and thus achieve overall improvements on learning performance. Due to limited quality or quality of prior knowledge, the agent should "smartly" inherit the most beneficial part and ignore harmful noise. We believe building a confidence mechanism that helps agents decide which policy to follow during its own learning would enhance the transfer efficiency. Considering the computation cost of Deep Learning, there are potential benefits to bootstrap in the early stage, by leveraging knowledge transfer.
Intelligent Robotics System
Intelligent Robotics System
Operating robotics system in real life is always difficult. We have been developing an intelligent bin management system for the harvest in orchards, funded by USDA. The purpose of this project is to build an intelligent multi-robot system to manage the usage of bins for harvest work in orchards. Our approach integrates the auto navigation of robots in an orchard environment and cooperation with human pickers. The value of this multi-robot bin managing system is in realizing the autonomous work of robots in tough outdoor environments and improving the harvest efficiency for the agriculture work.
Links:
Google Sites
Report abuse