Best AI Papers of the week (Dec 31 — Jan 4)

5 min readJan 9, 2019

In this post we’ll go over some of the most interesting papers of the previous week. This is no intended to be a comprehensive summary of the papers, but a quick overview of why each of these are interesting and some additional resources to learn more about the topic.

A Comprehensive Survey on Graph Neural Networks (https://arxiv.org/abs/1901.00596)

There are many times where data is represented as graphs with complex relationships. The complexity of graph data is a current challenge for many Machine Learning (ML) tasks. There are many studies that try to extend the Deep Learning (DL) approaches for graph data. This paper is a comprehensive overview of Graph Neural Networks, which were introduced in 2005. In the last 5 years, there has been a lot of research on graph convolutional networks (GCN), a set of approaches that generalize the convolution operator for graph data.

To understand what a GCN is, we suggest going over the following resources in the given order:

Deep Learning (DL) has achieved excellent results on Euclidean data. As an example, an image can be represented as a grid in the Euclidean space. Since graph data is irregular and nodes are interdependent, the assumptions of the DL algorithms don’t hold. There has been recent work trying to use ideas from convolutional and recurrent neural networks to design Graph Neural Networks.

One of the interesting contributions of this paper is the creation of a taxonomy of Graph Neural Networks. This taxonomy divides into Graph Convolution Network, which can be spatial-based (similar to how CNN filters go over an image) and spectral-based; Graph Autoencoders, which learn a low dimensional representation of a graph and reconstructs it; Graph Attention Network, which assigns attention weights to the nodes’ neighbors; Graph Generative Networks, which generate graphs given an observed set of graphs.

The paper also goes over some of the most common datasets and divides them into categories. Wrapping up, the paper covers different applications of GCNs like predicting semantic relationships, high-quality recommender system, and forecasting traffic speed.

Suggestions to learn about the topic

Improving fairness in machine learning systems: What do industry practitioners need? (https://arxiv.org/abs/1812.05239v1)

This paper is interesting since it gathers the result of 35 interviews and an anonymous survey of 267 ML practitioners in order to support fairer ML systems. During 2018, there has been a huge effort on developing algorithmic tools and methods that prevent ML systems that amplify social inequities and unfairness. An example is an automated hiring system which might perpetuate bias because that was the common patterns learned by the ML algorithms.

With the growth of ML usage in everyday-life systems (healthcare, education, social networks, jobs), there is also a bigger responsibility to prevent these systems to increase social inequality. As this field maturated, there have been mathematical definitions of fairness and algorithmic methods to detect discrimination in systems.

Let’s do a quick overview of the paper’s results:

Most of the respondents considered that the most important place to intervene is in the training datasets. Tools that facilitate communication between modelers and data collectors would be useful.
Careful test-set design could help detect potential fairness issues.
A challenge is to identify which subpopulations the teams need to consider when developing ML applications to ensure that there’s enough data to balance the datasets.
Sometimes the issues are detected when reported by customers’ complaints.
There’s a need for standard proactive auditing processes. Instead of responding to customers’ complaints, guidelines could help detect issues. Fairness metrics and KPI are a first step towards this.
Need for greater scalability. Current testing strategies are limited given the enormous space of possible fairness issues. Another issue is when gathering a diverse sample of user-study participants.

The paper gives some guidelines in how to address some of these issues. The first one is support to select a strategy. Depending on the type of problem, the team should focus in switching to a different model, augmenting training data, etc. To prevent wasting resources such as time, a strategy-methodology would be useful. When having to collect more data, it’s hard to determine how much additional data would be needed, normally having to rely on developers. A method to know how much additional data could be extremely useful. This paper is extremely interesting since it leaves room for a lot of potential interesting research that seeks to solve these issues.

Suggestions to learn more about this topic

Machine learning and AI research for Patient Benefit: 20 Critical Questions on Transparency, Replicability, Ethics and Effectiveness (https://arxiv.org/abs/1812.10404v1)

As Machine Learning and other statistical methods are being adopted in the health domain, there is an increasing awareness in the lack of transparency, replicability, exploration for potential ethical issues, and demonstrations of effectiveness. This paper is interesting since it provides 20 questions that should provide guidance towards following best practices in the entire project life-cycle.

The questions should help detecting potential issues, understanding what benefits using AI brings, trade-offs, transparency in data handling, validation methodologies, and model updating. Even if this paper is focused in healthcare ML, it’s worth it since many of these questions are useful for any kind of ML project.

DeepTraffic: Crowdsourced Hyperparameter Tuning of Deep Reinforcement Learning Systems for Multi-Agent Dense Traffic Navigation (https://arxiv.org/abs/1801.02805)

This is the second version of this paper, but I love this project, and I think it’s worth mentioning. DeepTraffic. DeepTraffic is an excellent traffic simulation where you can control some of the vehicles using a neural network. You can adapt the hyperparameters of this neural network and even submit your solution to the leaderboard. With almost 100 years of Reinforcement Learning simulation, there are many interesting results presented on this paper. There are 15 hyperparameters for the default DQN implementation that can be divided in three categories: input parameters that specify how the vehicles see the world, gradient descent parameters, and reinforcement learning parameters. Going over this paper is worth it since it has some insights on the results.

Other interesting papers from this week

An introduction to domain adaptation and transfer learning (https://arxiv.org/abs/1812.11806v1)
Accurate, Data-Efficient, Unconstrained Text Recognition with Convolutional Neural Networks (https://arxiv.org/abs/1812.11894v1)
A Theoretical Analysis of Deep Q-Learning (https://arxiv.org/abs/1901.00137v1)
Deep Reinforcement Learning for Multi-Agent Systems: A Review of Challenges, Solutions and Applications (https://arxiv.org/abs/1812.11794v1)

Thank you for reading!