Research

• Learning to reinforcement learn for Neural Architecture Search

  (2019) A novel approach to learning transferable RL policies through meta-learning to quickly design CNNs for image classification.

Open Source

• NasGym

  (2019) An OpenAI Gym environment for reproducible and comparable Neural Architecture Search with reinforcement learning. The default search space is inspired by BlockQNN and the performance estimation strategy is early-stop. It runs with TensorFlow and allows distributed training. The default datasets are CIFAR-10 and the meta-dataset. The code allows to plug-in different search spaces, performance estimation strategies, and datasets.

• The meta-A2C learner

  (2019) A TensorFlow implementation of the meta-A2C algorithm proposed in DeepMind's paper "Learning to reinforcement learn". The code is written on top of the OpenAI baselines to enable straightforward integration with Gym environments.

• AutoML assistant

  (2018) A python package that guides Data Scientists to select the best scikit-learn model for classification and regression tasks. The assistance includes fully-automated pipeline generation using the TPOT, model recommendation with meta-learning, and hyperparameter tuning with bayesian optimization (SMAC).