Export
Tomohiro Tanaka
· 2 previous affiliations
- NTT (Japan)Since 2017
- Researcher · Since 2017
- Tokyo Institute of Technology2015–2017
- MS student · 2015–2017
Research Overview
Tomohiro Tanaka focuses on automated machine learning, large vocabulary speech recognition, and evolutionary algorithms. His trajectory centers on applying covariance matrix adaptation evolution strategies to optimize the architecture and hyper-parameters of complex neural network systems. The most recent evidence of his work emphasizes multi-objective optimization for high-performance audio and language processing models.
Research Trajectory
Between 2015 and 2019, Tanaka's research primarily addressed the significant manual effort required to tune deep neural network (DNN) and hidden Markov model (HMM) architectures for speech recognition. His work systematically applied evolutionary algorithms to automate the discovery of optimal network structures and training configurations, treating hyper-parameter tuning as a rigorous optimization problem rather than a manual heuristic process.
In 2015, Tanaka was the second author on Automation of system building for state-of-the-art large vocabulary speech recognition using evolution strategy. This research introduced a method to tune DNN-HMM based large vocabulary speech recognition systems using the covariance matrix adaptation evolution strategy (CMA-ES). By framing the tuning process as a multi-objective Pareto optimization problem, the work successfully balanced the competing goals of achieving high recognition accuracy and maintaining a compact model size. The approach also demonstrated practical scalability, proving to be efficiently parallelizable for cloud computing environments.
Building on this foundation, Tanaka led research as the first author on Automated structure discovery and parameter tuning of neural network language model based on evolution strategy in 2016. This paper extended the application of CMA-ES to long short-term memory (LSTM) recurrent neural networks. Rather than solely tuning continuous hyper-parameters, this work allowed for the flexible optimization of various meta-parameters, including layer-wise unit types. The automated configuration process successfully identified network structures that improved overall speech recognition performance without requiring exhaustive manual trial and error by human experts.
Tanaka's focus on automated system development culminated in a comprehensive 2019 journal publication, Evolution-Strategy-Based Automation of System Development for High-Performance Speech Recognition, where he served as second author. This work consolidated earlier findings into a unified framework for optimizing the numerous meta-parameters that dictate the design and training setups of entire large vocabulary speech recognition pipelines, ensuring systems were optimized for both low word error rates and structural compactness.
While Tanaka's core publications focus on speech and language model optimization, his broader research profile intersects with contemporary challenges in model tuning and adaptation. The underlying principles of his work—efficiently navigating complex parameter spaces to adapt large models without degrading performance—align conceptually with modern machine learning challenges, such as the research direction Synchronized Cross-Modal Alignment for Vision-Language Model Adaptation and the broader area of Manifold Preservation for Robust Foundation Model Adaptation.
Current Directions
The most recent paper in the provided evidence is from 2019. Based on this work, Tanaka's latest documented research investigates the end-to-end automation of system development for high-performance speech recognition. The primary open problem addressed in this recent phase is the elimination of manual expert tuning in favor of multi-objective evolutionary optimization. By leveraging CMA-ES, this direction seeks to systematically generate acoustic and language models that are strictly optimized for both minimal word error rates and computational efficiency, providing a scalable blueprint for deploying complex neural architectures.
Profile updated
Loading new papers…
Loading research neighbors…
See something wrong? Report a mistake