Cutting Edge
IAAI-20
Real-world application of AI technology
Jul 17, 2020
The IAAI-20 (Innovative Applications for Artificial Intelligence) is a sister conference within the larger AAAI-20 held every year focusing on applications of AI in the real-world. The conference combines technical papers, invited talks and panel discussions to showcase and discuss new ways of applying AI to real-world problems.
Sony is interested in keeping up with the activities of its peers and learning what problems are gaining focus in academic and industrial communities. Although the conference covers a variety of fields broadly classified under 9 tracks, we briefly present some works related to Vision, NLP and Recommendation Systems.
Calorie Estimation In a Real-World Recipe Service
As Japan’s largest recipe sharing service, a major objective is to automate the calculation of caloric value per serving in any recipe submitted by its users. This was reduced to solving two sub-problems; of identifying each ingredient and estimating the number of serving the recipe makes. An encoder-decoder model was proposed for each sub-problem and trained on thousands of annotated examples to give significantly better accuracy than traditional regular-expression based methods. This solution is currently deployed on their service.
It was interesting to see how much colloquial descriptions of ingredients in Japanese can diverge from their canonical definitions. Future work is said to make use of dish characteristics to improve estimation of the number of servings.
Improving lives of Indebted Farmers using Deep Learning: Predicting Agricultural Produce Prices Using Convolutional Neural Networks
An academic work aimed to reduce farmer suicides in India by proposing crops that will yield high market prices using past Price & Volume information. Data was collected using publicly available Indian Govt. databases and Google Maps API to train a Wide & Deep CNN structure that predicted future prices for the crops it was trained over. It outperformed related works with a 25% lesser CoV.
The work claims to be the first to explicitly model Spatio-Temporal dependence on a price prediction model. The authors are in collaboration with a non-profit organization to deploy it within select states in the country.
Understanding Chat Messages for Sticker Recommendation in Messaging Apps
With a majority of its 100 million active user-base consisting of Indians, the authors present some of their unique issues and solutions when developing a Sticker Recommendation solution for their platform. The solution proposed predicted the user’s next message and substitute it with an appropriate sticker from the platform database.
This was specifically interesting because in the Indian context, texts are usually transliterated from native Indian languages to roman letters and mixed with English words/phrases as well. They learnt dense representations of chat texts to cluster and map them. The solution was optimized for speed and memory so that it works on low-end smartphones under limited network bandwidths a common problem in the Indian markets.
Summary
The IAAI focuses on successful applications and novel uses of AI in the real world. Many submissions this year were aimed at reducing societal problems and minimize monetary/human loss.
Although most submissions did not extend the State-of-the-art in their respective areas, they usually represent first successful applications of algorithms that have only shown experimental viability on real-world scenarios and how modifications were made to deal with specific issues that arose in each application.
The conference serves as a platform for small and large businesses to showcase and appeal their work using AI. Besides serving as a stage for healthy discussions, many groups seek to network and form potential collaborations in the future.
Reporter
Name : Akshat Kadam
Education : Electrical
Current job : Solving problems relevant to the Indian markets
Job Category : Software, Signaling, Information Processing
