plug and play language model
In plain wordsAn early way to steer what a language model writes — its topic or mood — by letting a small add-on classifier nudge the model's internal state at each word, without retraining the model.
A method for controllable text generation that leaves a pretrained language modellarge language model, n.: A computer program trained on enormous amounts of text to predict the words that come next. At vast scale, that simple skill lets it chat, answer questions, translate, summarize and write code. untouched and steers it during decoding with a small, separately supplied scorer for the desired attribute: a list of topic words, or a one-layer sentiment classifier on the model's hidden states. At each generation step, gradients of the attribute model's log-likelihood are propagated back into the language model's cached key–value activations, which are nudged toward the attribute for a few iterations, while a KL-divergenceKullback–Leibler divergence, n.: A score for the cost of believing the wrong odds: on average, how much more surprised you will be by real outcomes than someone who knew the true probabilities. It is zero only when your odds are exactly right. penalty and a fusion with the unmodified distribution keep the text fluent. It steered GPT-2GPT-2, prop. n.: A 2019 OpenAI language model that wrote surprisingly fluent paragraphs. OpenAI at first withheld the full version, saying it could be misused for fake news, which made it famous. toward topics and sentiments and reduced toxicity without fine-tuning, but the per-token gradient updates made generation several times slower. PPLM is a direct ancestor of activation steeringactivation steering, n.: Changing an AI model's behavior by adjusting the internal numbers it computes while it runs — like turning a hidden dial — instead of rewording the prompt or retraining the model., which replaces this optimization with fixed vectors.
With the 'space' bag of words, PPLM turned a prompt about potatoes into a story about orbiting satellites.
Origin
Introduced by Sumanth Dathathri, Andrea Madotto, Janice Lan, Jane Hung, Eric Frank, Piero Molino, Jason Yosinski and Rosanne Liu, mainly of Uber AI, in 'Plug and Play Language Models: A Simple Approach to Controlled Text Generation' (arXiv, December 2019; ICLR 2020). The name echoes the image-generation 'Plug & Play Generative Networks' (Anh Nguyen and colleagues, 2017).
Related
- activation steeringactivation steering, n.: Changing an AI model's behavior by adjusting the internal numbers it computes while it runs — like turning a hidden dial — instead of rewording the prompt or retraining the model.
- steeringsteering, n.: Guiding what an AI model does — its topic, tone or behavior — while it runs, by changing its instructions, its internal signals or how its next word is chosen, rather than by retraining it.
- GPT-2GPT-2, prop. n.: A 2019 OpenAI language model that wrote surprisingly fluent paragraphs. OpenAI at first withheld the full version, saying it could be misused for fake news, which made it famous.
- steering vectorsteering vector, n.: A direction inside an AI model's internal activity that, when added while the model runs, nudges its behavior — say, more upbeat, more honest or fixated on a topic — without retraining it.
- classifier-free guidanceclassifier-free guidance, n.: A setting that makes an image generator follow the prompt more strictly by comparing what it would draw with and without the prompt, then exaggerating the difference.
Sources
- Dathathri, S. et al. (2020). Plug and Play Language Models: A Simple Approach to Controlled Text Generation. ICLR 2020.
Mentioned in
activation steeringactivation steering, n.: Changing an AI model's behavior by adjusting the internal numbers it computes while it runs — like turning a hidden dial — instead of rewording the prompt or retraining the model., steeringsteering, n.: Guiding what an AI model does — its topic, tone or behavior — while it runs, by changing its instructions, its internal signals or how its next word is chosen, rather than by retraining it.