Li Mingda Email & Phone Number

Seattle, Wa, Us

In Amazon Alexa AI: Deep NLU (Natural Langauge Understanding) team, work as an Applied Scientist.Publications as follows:[1] Multi-task Learning of Spoken Language Understanding by Integrating N-Best Hypotheses with Hierarchical Attention; [COLING 20'] Mingda Li, Xinyue Liu, Weitong Ruan, Luca Soldaini, Wael Hamza, Chengwei Su https://www.aclweb.org/anthology/2020.coling-industry.11/[2] Improving spoken language understanding by exploiting asr n-best hypotheses Mingda Li, Weitong Ruan, Xinyue Liu, Luca Soldaini, Wael Hamza, Chengwei Su https://arxiv.org/abs/2001.05284[3]ASR n-best fusion nets [ICASSP 21'] Xinyue Liu, Mingda Li, Luoxin Chen, Prashan Wanigasekara, Weitong Ruan, Haidar Khan, Wael Hamza, Chengwei Su https://www.amazon.science/publications/asr-n-best-fusion-nets

Seattle, Wa, Us

Work in Alexa AI NLU group. Use deep learning models (transformer, bi-lstm) to strengthen the Natural Language Understanding (NLU) tasks (Domain classification, Intent classification). One paper as following about integrating ASR N bests to enhance Spoken Language Understanding:https://arxiv.org/abs/2001.05284

Seattle, Wa, Us

I worked in the AWS Redshift Query Processing Team. One part of my internship was to utilize machine learning algorithms to predict query resource requirement and design a queuing policy to manage workloads based on prediction result. I also worked on Short Query Acceleration (SQA) to persist timeout and the machine learning classifier model to seamlessly accelerate short queries after a cluster is rebooted. This was announced by the Redshift service in September 2018.

Princeton, Nj, Us

Work on an artificial way for IP spoofing defense. During the process, we developed many different models (linear and non-linear) making the final system accurate, efficient and open for new IPs. The project is a combination of network security and machine learning. It win a best paper award : https://dl.acm.org/citation.cfm?id=3229609 Three patents generated (listed here):https://scholar.google.com/citations?hl=en&user=j0vd1oYAAAAJ&view_op=list_works&sortby=pubdateWin the best paper awards of ACM Sigcomm Bigdama 2018!!

Los Angeles, Ca, Us

Teaching Assistant of CS 33 about Computer Organization.

Los Angeles, Ca, Us

Teaching Assistant of Course CS 143: Database Systems.

San Diego, California, Us

In Optimizer Group (Query Optimization) and work on Multiple Join Optimization. Use look ahead algorithm to help optimizer peel the next n joins for execution and collect stats feedback. This will help the optimizer generate a more accurate plan in the following process. There are some challenges including converting the query according to the n joins each time we peel out such that the executer could finish the job in our order, deciding which n joins to peel every time, set the amount of n and the stop point of the algorithm. During the process, I learned the whole process of optimization job and can get the join plan to analyze if it is better and why. This is mainly finished by Python on PyCharm (I use a Graph structure helping me to convert query according to the peeling result).

Los Angeles, Ca, Us

Combine Spark (powerful machine learning and graph libraries and fast execution), Apache AsterixDB (fast ingestion and indexing of semi-structured data and efficient answering of geo-spatial and fuzzy text queries) and Zeppelin (interactive data analytics) together to support big data analytics in cluster. Now, I am converting Spark SQL queries to AsterixDB AQL queries to reduce the amount of data transferred between them. This is a starting point of a Hybrid Database for my second year.Choose best join plan for the multiple join in Hyracks (query execution layer of AsterixDB) and Spark through building Bloom Filter to filter items and collecting stats information meanwhile