A widespread strategy to obtain a language model that performs well on a target domain is to finetune a pretrained model to perform unsupervised next-token prediction on data from that target domain. Finetuning presents two challenges: (i) if the amount of target data is limited, as in most practical applications, the model will quickly overfit, and (ii) the model will drift away from the original model, forgetting the pretraining data and the generic knowledge that comes with it. Our goal is to derive scaling laws that quantify these two phenomena for various target domains, amounts of available target data, and model scales. We measure the efficiency of injecting pretraining data into the finetuning data mixture to avoid forgetting and mitigate overfitting. A key practical takeaway from our study is that injecting as little as $1%$ of pretraining data in the finetuning data mixture prevents the model from forgetting the pretraining set.

The growing use of generative models in daily life calls for efficient mechanisms to control their generation, to e.g. produce safe content or provide users with tools to explore style changes. Ideally, such mechanisms should require low volume of unpaired data (\ie without explicit preference), and should be cheap, both at train and inference time, while preserving output quality. Recent research has shown that such mechanisms can be obtained by intervening exclusively on model activations, with the goal of correcting distributional differences between activations seen when using prompts from a source vs. a target set (e.g. toxic and non-toxic sentences). While cheap, these fast methods are inherently crude: their maps are tuned locally, not accounting for their impact on downstream layers, resulting in interventions that cause unintended shifts when used out-of-sample. We propose in this work linear end-to-end activation steering (LinEAS), an approach trained with a global loss that accounts simultaneously for all layer-wise distributional shifts. In addition to being more robust, the loss used to train LinEAS can be regularized with sparsifying norms, which can automatically carry out neuron selection. LinEAS only requires a handful of unpaired samples to be effective, and beats similar baselines on toxicity mitigation in language models, becoming competitive with oracle-dependent methods that have access to strong supervision. LinEAS is modality-agnostic and we empirically find that it outperforms existing activation steering methods at mitigating and including new concepts at the output of single-step text-to-image generation models.

Human lexicons contain many different words that speakers can use to refer to the same object, e.g., purple or magenta for the same shade of color.On the one hand, studies on language use have explored how speakers adapt their referring expressions to successfully communicate in context, without focusing on properties of the lexical system.On the other hand, studies in language evolution have discussed how competing pressures for informativeness and simplicity shape lexical systems, without tackling in-context communication.We aim at bridging the gap between these traditions, and explore why a soft mapping between referents and words is a good solution for communication, by taking into account both in-context communication and the structure of the lexicon.We propose a simple measure of informativeness for words and lexical systems, grounded in a visual space, and analyze color naming data for English and Mandarin Chinese.We conclude that optimal lexical systems are those where multiple words can apply to the same referent, conveying different amounts of information.Such systems allow speakers to maximize communication accuracy and minimize the amount of information they convey when communicating about referents in contexts.

Human languages support both semantic categorization and local pragmatic interactions that require context-sensitive reasoning about meaning. While semantics and pragmatics are two fundamental aspects of language, they are typically studied independently and their co-evolution is largely under-explored. Here, we aim to bridge this gap by studying how a shared lexicon may emerge from local pragmatic interactions. To this end, we extend a recent information-theoretic framework for emergent communication in artificial agents, which integrates utility maximization, associated with pragmatics, with general communicative constraints that are believed to shape human semantic systems. Specifically, we show how to adapt this framework to train agents via unsupervised pragmatic interactions, and then evaluate their emergent lexical semantics. We test this approach in a rich visual domain of naturalistic images, and find that key human-like properties of the lexicon emerge when agents are guided by both context-specific utility and general communicative pressures, suggesting that both aspects are crucial for understanding how language may evolve in humans and in artificial agents.

Neural captioners are typically trained to mimic human-generated references without optimizing for any specific communication goal, leading to problems such as the generation of vague captions. In this paper, we show that fine-tuning an out-of-the-box neural captioner with a self-supervised discriminative communication objective helps to recover a plain, visually descriptive language that is more informative about image contents. Given a target image, the system must learn to produce a description that enables an out-of-the-box text-conditioned image retriever to iden-tify such image among a set of candidates. We experiment with the popular ClipCap captioner; also replicating the main results with BLIP. In terms of similarity to ground-truth human descriptions, the captions emerging from dis-criminative finetuning lag slightly behind those generated by the non-finetuned model, when the latter is trained and tested on the same caption dataset. However, when the model is used without further tuning to generate captions for out-of-domain datasets, our discriminatively-finetuned captioner generates descriptions that resemble human ref-erences more than those produced by the same captioner wihtout finetuning. We further show that, on the Concep-tual Captions dataset, discriminatively finetuned captions are more helpful than either vanilla ClipCap captions or ground-truth captions for human annotators tasked with an image discrimination task.11Our code is available at https://github.com/facebookresearch/EGG/tree/main/egg/zoo/discriminative_captioner.

We compare the 0-shot performance of a neural caption-based image retriever when given as input either human-produced captions or captions generated by a neural captioner.We conduct this comparison on the recently introduced IMAGECODE data-set (Krojer et al., 2022), which contains hard distractors nearly identical to the images to be retrieved.We find that the neural retriever has much higher performance when fed neural rather than human captions, despite the fact that the former, unlike the latter, were generated without awareness of the distractors that make the task hard.Even more remarkably, when the same neural captions are given to human subjects, their retrieval performance is almost at chance level.Our results thus add to the growing body of evidence that, even when the "language" of neural models resembles English, this superficial resemblance might be deeply misleading.