These are this year's most exciting pieces of AI study. It brings together advances in AI and data science. It is set up in order of time and has a link to a longer piece.

Game Action Modeling for Fine-Grained Analyses of Player Behavior in Multi-player Card Games

The researchers show a deep learning system for modelling game actions. It makes it possible to look closely at how players act. They make CNN-based supervised models that learn from skilled players how to make essential game decisions, and they use these models to figure out things about players in the system, like how long they stay, how much they pay, and so on. The researchers show that a carefully designed input format that shows the game state and history as a multidimensional picture and a custom architecture help the model learn the game's strategies well. 

The self-play simulation also improves it, letting it "look ahead" and better understand the game state. This information is used in a new loss function. Next, the experts show that comparing the players to these models is a constructive way to determine how players can be engaged and make money. They also use the model to determine the situations in which players make mistakes. They then use this information to help players improve their skills.

UNMASQUE: A Hidden SQL Query Extractor

When given a database instance and a result, query reverse engineering determines which SQL queries could be run on the instance to get this result. A version of this problem comes up when a ground truth is also available but is hidden in a computer program. In this demo, the researchers show UN-MASQUE, an extraction method that can find a large class of these hidden queries with high accuracy. Its design is based on the fact that the extraction doesn't hurt the application. 

In particular, it only looks at the results of application runs on databases with a mix of data mutation and data generation methods. It makes it independent of the platform. Also, powerful improvements, like reducing the size of the database to just a few rows, are added to reduce the extraction overheads as much as possible. The video shows how both declarative and imperative applications can use these features.

Learning Colour Representations of Search Queries

Image search engines rely on well-designed ranking criteria that capture various aspects of the material's semantics and historical popularity. The researchers investigate the role of colour in this relevant matching procedure in this paper. Their research is prompted by the discovery that many user inquiries have an inherent colour connected with them. While some searches have explicit colour mentions (for example, 'black automobile' and 'yellow daisies'), others have implicit colour conceptions (for example, 'sky' and 'grass'). 

Furthermore, grounding inquiries in colour is a distribution in colour space rather than a mapping to a single colour. For example, a search for 'trees' has a bimodal distribution centred on green and brown colours. The researchers use historical clickthrough data to generate a colour representation for search queries and propose a recurrent neural network design to encode previously unknown requests into colour space. They also demonstrate how this embedding may be learned alongside a cross-modal relevance ranker from impression logs containing a fraction of the result images clicked. The researchers show that using a query-image colour distance feature improves ranker performance as evaluated by users' preferences for clicked versus skipped images.

Disclaimer: The above content is published to promote India's AI ecosystem and educate the public about AI and its developments. However, it should be noted that INDIAai and MeitY do not endorse or have affiliations with the products, startups, and organizations mentioned in the article. Readers are advised to conduct their research and due diligence before engaging with any entities mentioned.

Image source: Unsplash