Experiments with OpenAI’s ‘preference learning’ approach, which trains a NN to predict global quality of datapoints, and then uses reinforcement learning to optimize that directly, rather than proxies. I am unable to improve quality, perhaps due to too-few ratings.
Standard language generation neural network models, like GPT-2, are trained via likelihood training to imitate human text corpuses. Generated text suffers from persistent flaws like repetition, due to myopic generation word-by-word, and cannot improve on the training data because they are trained to predict ‘realistic’ completions of the training data.
A proposed alternative is to use reinforcement learning to train the NNs, to encourage global properties like coherence & lack of repetition, and potentially improve over the original corpus’s average quality. Preference learning trains a reward function on human ratings, and uses that as the ‘environment’ for a blackbox DRL algorithm like PPO.
OpenAI released a codebase implementing this dual-model preference learning approach for textual generation, based on GPT-2. Having previously used GPT-2 for poetry & music generation, I experimented with GPT-2 preference learning for unconditional music and poetry generation.
I found that preference learning seemed to work better for music than poetry, and seemed to reduce the presence of repetition artifacts, but the results, at n ≈ 7,400 ratings compiled over 23 iterations of training+sampling November 2019–January 2020, are not dramatically better than alternative improvements like scaling up models or more thorough data-cleaning or more stringent sample curation. My blind ratings using n ≈ 200 comparisons showed no large advantage for the RL-tuned samples (winning only 93 of 210 comparisons, or 46%).
This may be due to insufficient ratings, bad hyperparameters, or not using samples generated with common prefixes, but I suspect it’s the former, as some NLP tasks in et al 2019 required up to 60k ratings for good performance, and the reward model appeared to achieve poor performance & succumb to adversarial examples easily.
Working with it, I suspect that preference learning is unnecessarily sample-inefficient & data-inefficient, and that the blackbox reinforcement learning approach is inferior to directly using the reward model to optimize text samples, and propose two major architectural overhauls: have the reward model directly model the implied ranking of every datapoint, and drop the agent model entirely in favor of backprop-powered gradient ascent which optimizes sequences to maximize the reward model’s output.