These are the year's most intriguing AI research publications. It integrates innovations in artificial intelligence (AI) and data science. It is chronologically organized and includes a link to a lengthier article.

Near-optimal sample complexity bounds for learning Latent k-polytopes and applications to Ad-Mixtures

An active study area is finding the best limits on the sample complexity of models with latent variables. Recently, these kinds of limitations were found for Mixtures of Gaussians. However, no such limits have been found for Ad-mixtures, a type of Mixture distribution.

The researchers focus on Mixed Membership Stochastic Block Models (MMSB) to show that the method works in various situations. They show that if an MMSB has k blocks, the sample complexity is O∗(k2), the same as the usual assumption.

The Optimal Sample complexity of LKP shows that it changes linearly with the number of factors. The universality of LKP and the tools made point to a simple way to get Sample complexity estimates for Ad-mixture models. You can get the same outcomes from other models, like Dirichlet Simplex Nest (Yurochkin et al., 2019).

Table Structure Recognition using Top-Down and Bottom-Up Cues

Tables are structured items in document images that hold a lot of information. Much work has gone into finding tables as graphic items in document images, but more needs to be done to recognize table structures. A lot of the research that has already been done on structure recognition relies on either optical character recognition (OCR) models to get low-level layout features from images or extracting meta-features from PDF files. However, these methods only work well in some cases because they need meta-features or because the OCR needs to be corrected when there are significant differences in how tables are laid out and how text is organized. The experts are primarily interested in tables with complicated structures, lots of information, and different layouts that don't need meta-features or OCR.

The researchers show a way to recognize the structure of a table that mixes modules for finding cells and interacting with them to figure out where the cells are and guess how they will be connected to other cells in rows and columns. To find cells, they add structural restrictions as extra differential components to the loss function. In their work, they took table structure recognition in a new way by using top-down (table cells detection) and bottom-up (structure recognition) cues to understand the tables visually.

Unsupervised Domain Adaptation for Semantic Segmentation of NIR Images through Generative Latent Search

Separating the pixels that belong to human skin is an essential first step in many situations, from spying to figuring out heart rate from remote photoplethysmography. However, the research that has been done so far only looks at the issue in the visible part of the EM spectrum. It means they can only be used in places with light, where the application is most important. Researchers are looking into how to solve the problem of skin segmentation from near-infrared pictures to avoid this problem. However, the most up-to-date segmentation methods based on deep learning need a lot of labelled data, which is unavailable in this situation. 

To solve this problem, they turned it into a target-independent Unsupervised Domain Adaptation (UDA) problem. To do this, they used data from the visible range's red channel to create a skin segmentation method for NIR images. The researchers suggest a way to segment images that don't depend on the target. They say that the "nearest clone" of the target image in the source domain should be looked for and used as a stand-in in a segmentation network that has only been trained on the source domain. The existence of the "nearest clone" is proven, and a way to find it is suggested. It would be to use an optimization technique over the latent space of a Deep generative model. 

The researchers show that the suggested method for NIR skin segmentation works better than the most recent UDA segmentation methods on two brand-new NIR skin segmentation datasets, even though they need access to the target NIR data. On top of that, they show the best results for adapting from Synthia to Cityscapes.

Image source: Unsplash