The IEEE Signal Processing Society, the leading professional association for signal processing and data science engineers, recently awarded Mike Schuster, Head of Two Sigma’s AI Core Team and his co-author, Kuldip K. Paliwal, the 2021 IEEE SPS Sustained Impact Paper Award for their paper “Bidirectional Recurrent Neural Networks,” published in 1997 in the IEEE Transactions on Signal Processing.
Bidirectional Recurrent Neural Networks
The paper, which has been cited more than 7,000 times, investigates different artificial neural network (ANN) structures for incorporating temporal dynamics of sequences. The authors conduct a variety of experiments using both artificial and real-world data, and show the superiority of recurrent neural networks (RNNs) over the other structures. Finally, they highlight some of the limitations of RNNs and propose a modified version of an RNN called a bidirectional recurrent neural network (BRNN), which overcomes these limitations. The paper also shows how to use a modified BRNN to maximize the complete conditional probability of a symbol sequence with matching input/output lengths, an early form of today’s more general Sequence-to-Sequence models.
Bidirectional recurrent neural networks have become a fundamental building block of many neural network systems that deal with sequential data. Today they can be found in state-of-the-art systems for speech recognition, speech synthesis, machine translation, part-of-speech tagging, protein prediction, hand-writing recognition, financial time-series analysis, and many more.
Read the paper here.