While large-scale language models (LMs) are able to imitate the distribution of natural language well enough to generate realistic text, it is difficult to control which regions of the distribution they generate. This is especially problematic because datasets used for training large LMs usually contain significant toxicity, hate, bias, and negativity. One promising approach to address this is to use discriminators to guide decoding from LMs, but existing methods for this are too slow to be useful in practice for many applications. We present GeDi as a significantly more efficient discriminator-based approach for guiding decoding. GeDi guides generation at each step by computing classification probabilities for all possible next tokens via Bayes rule by normalizing over two class-conditional distributions; one conditioned on the desired attribute, or control code, and another conditioned on the undesired attribute, or anti control code. We find that GeDi gives controllability on par with or better than previous controllable generation methods. GeDi results in significantly faster generation speeds than the only previous method that achieved comparable controllability in our experiments. We also show that GeDi can make GPT-2 and GPT-3 significantly less toxic while maintaining linguistic fluency, without sacrificing significantly on generation speed. Lastly, we find training GeDi on only three topics allows us to controllably generate new topics zero-shot from just a keyword.

No abstract

Class-conditional language models (CC-LMs) can be used to generate natural language with specific attributes, such as style or sentiment, by conditioning on an attribute label, or control code. However, we find that these models struggle to control generation when applied to out-of-domain prompts or unseen control codes. To overcome these limitations, we propose generative discriminator (GeDi) guided contrastive generation, which uses CC-LMs as generative discriminators (GeDis) to efficiently guide generation from a (potentially much larger) LM towards a desired attribute. In our human evaluation experiments, we show that GeDis trained for sentiment control on movie reviews are able to control the tone of book text. We also demonstrate that GeDis are able to detoxify generation and control topic while maintaining the same level of linguistic acceptability as direct generation from GPT-2 (1.5B parameters). Lastly, we show that a GeDi trained on only 4 topics can generalize to new control codes from word embeddings, allowing it to guide generation towards wide array of topics. * Equal Contribution † Work performed while at Salesforce Research 1 pronounced "Jedi" 2 Code will be released at https://github.com/salesforce/GeDi

The convergence rate and final performance of common deep learning models have significantly benefited from heuristics such as learning rate schedules, knowledge distillation, skip connections, and normalization layers. In the absence of theoretical underpinnings, controlled experiments aimed at explaining these strategies can aid our understanding of deep learning landscapes and the training dynamics. Existing approaches for empirical analysis rely on tools of linear interpolation and visualizations with dimensionality reduction, each with their limitations. Instead, we revisit such analysis of heuristics through the lens of recently proposed methods for loss surface and representation analysis, viz., mode connectivity and canonical correlation analysis (CCA), and hypothesize reasons for the success of the heuristics. In particular, we explore knowledge distillation and learning rate heuristics of (cosine) restarts and warmup using mode connectivity and CCA. Our empirical analysis suggests that: (a) the reasons often quoted for the success of cosine annealing are not evidenced in practice; (b) that the effect of learning rate warmup is to prevent the deeper layers from creating training instability; and (c) that the latent knowledge shared by the teacher is primarily disbursed to the deeper layers.

No abstract