The thirty-fifth AAAI Conference on Artificial Intelligence (AAAI-21) was hosted virtually this year due to the COVID-19 pandemic restrictions. However, there was no dearth of participation in the event from across the world. This year, 1,692 research papers were selected from more than 9,000 papers submissions. 

The AAAI committee has declared the following paper entries as the best of the lot:

Informer: Ahead of Efficient Transformer for Long Sequence time-series Forecasting 

 A Long Sequence Time-series Forecasting (LSTF) helps capture precise long-range dependency between output and inputs such as electricity consumption. The paper presents Informer, created by researchers at UC Berkeley which has three unique characteristics as compared to other transformer-based models for LSTF:

• a ProbSparse Self-attention mechanism, which achieves O(LlogL) in time complexity and memory usage, and has comparable performance on sequences' dependency alignment. 
• the self-attention distilling highlights dominating attention by halving cascading layer input, and efficiently handles extreme long input sequences. 
• the generative style decoder, while conceptually simple, predicts the long time-series sequences at one forward operation rather than a step-by-step way, which drastically improves the inference speed of long-sequence predictions. 

Exploration-Exploitation in MAL: Game Theory together with Catastrophe Theory

Exploration searches the whole sample space for input while exploitation literally exploits promising areas found from the exploration. This approach is a powerful tool when used for multi-agent learning (MAL). Singapore-based researchers from the Singapore Univerisity of Technology presented a paper that presented their study of an adaptation of stateless Q-learning with Boltzmann Q-learning or smooth Q-learning (SQL). The researchers, through their paper "provide a formal theoretical treatment of how tuning the exploration parameter can provably lead to equilibrium selection with both positive as well as negative (and potentially unbounded) effects to system performance."

Reducing political biases in GPT-2

A group of researchers from Dartmouth College, the University of Texas and ProtagoLabs submitted a paper on their work to describe metrics for quantifying political prejudice in GPT-2 so that they can reduce political biases in the generated text by using reinforcement learning (RL). The RL framework used rewards for word embeddings without having access to the training data or requiring the model to be retrained.