OpenAI Scholars 2018 Final Projects

OpenAI has released a summary of the final projects completed by the 2018 Scholars cohort. These projects demonstrate the application of deep learning, reinforcement learning, and generative models across creative arts, physical simulation, and linguistic analysis.

Generative Music and Audio

Several scholars focused on using language modeling and generative architectures to create music and music-related content.

  • Classical Music Generation: Christine Payne developed a system that adapts language modeling techniques to model classical music using both notewise and chordwise approaches. The system utilizes parallel character and word-level language models to generate solo piano or piano and violin chamber music, with the ability to train on specific composers or styles. The project includes a music critic network to distinguish real from fake samples and a network to identify the composer.
  • Music Commentary Generation: Nadja Rhodes created "Deephypebot," a language model trained on human music writing from the web. The model is conditioned on music attributes and includes a specific training step to encourage the "flowery" descriptive writing style typical of music journalism.

Natural Language Processing and Semantics

Research in the 2018 cohort focused on improving how models represent meaning and semantic relationships in text.

  • Semantic Trees in LSTMs: Munashe Shumba explored the use of semantic trees within Long Short-Term Memory (LSTM) networks to better represent the relationships between entities within a sentence.
  • Semantic Textual Similarity: Dolapo Martins built a model for the SemEval STS Task, Track 4, experimenting with various architectures including LSTMs, RNNs, and the Transformer architecture to optimize performance.

Creative Arts and Visual Generation

Scholars applied generative adversarial networks (GANs) and variational autoencoders (VAEs) to visual art and emotional expression.

  • Art Composition Attributes: Holly Grimm developed an Art Composition Attributes network using a pretrained ResNet50 network fine-tuned on eight art composition attributes. This was integrated into a CycleGAN to generate art compositions based on target values for those attributes.
  • Emotional Landscapes: Hannah Davis experimented with generating emotional landscapes using a GAN, a conditional VAE, and a multi-scale VAE, mapping specific emotion classes such as joy, anticipation, fear, sadness, surprise, trust, disgust, and anger.

Physical Simulation and Grounded Learning

Two projects focused on the intersection of AI and the physical or environmental world.

  • Intuitive Physics: Ifu Aniemeka built a network capable of learning the rules of object movement in space through observation alone, without explicit definitions of physics concepts like friction, force, or momentum.
  • Grounded Language Learning: Sophia Arakelyan developed a project combining NLP and reinforcement learning (RL) where an agent achieves target cells in a Gridworld environment based on received commands, serving as a step toward grounded language learning.

Technical Skill Acquisition

The scholars reported significant growth in their technical fluency with industry-standard frameworks and methodologies. Key areas of development included:

  • Frameworks: Increased proficiency in TensorFlow, PyTorch, and Keras.
  • Architectures: Hands-on implementation of Recurrent Neural Networks (RNNs), LSTMs, Generative Adversarial Networks (GANs), and Variational Autoencoders (VAEs).
  • Methodologies: Experience with data generation, feature engineering, and the application of reinforcement learning algorithms.

Sources