Making language models bigger does not inherently make them better at following a user's intent. For example, large language models can generate outputs that are untruthful, toxic, or simply not helpful to the user. In other words, these models are not aligned with their users. In this paper, we show an avenue for aligning language models with user intent on a wide range of tasks by fine-tuning with human feedback. Starting with a set of labeler-written prompts and prompts submitted through a language model API, we collect a dataset of labeler demonstrations of the desired model behavior, which we use to fine-tune GPT-3 using supervised learning. We then collect a dataset of rankings of model outputs, which we use to further fine-tune this supervised model using reinforcement learning from human feedback. We call the resulting models InstructGPT. In human evaluations on our prompt distribution, outputs from the 1.3B parameter InstructGPT model are preferred to outputs from the 175B GPT-3, despite having 100x fewer parameters. Moreover, InstructGPT models show improvements in truthfulness and reductions in toxic output generation while having minimal performance regressions on public NLP datasets. Even though InstructGPT still makes simple mistakes, our results show that fine-tuning with human feedback is a promising direction for aligning language models with human intent.

To Perform reliably and consistently over sustained periods of time, large-scale automation critically relies on computer simulation. Simulation allows us and supervisory AI to effectively design, validate, and continuously improve complex processes, and helps practitioners to gain insight into the operation and justify future investments. While numerous successful applications of simulation in industry exist, such as circuit simulation, finite element methods, and computeraided design (CAD), state-of-the-art simulators fall short of accurately modeling physical phenomena, such as friction, impact, and deformation.

Deep learning-based robotic grasping has made significant progress thanks to algorithmic improvements and increased data availability. However, state-of-the-art models are often trained on as few as hundreds or thousands of unique object instances, and as a result generalization can be a challenge. In this work, we explore a novel data generation pipeline for training a deep neural network to perform grasp planning that applies the idea of domain randomization to object synthesis. We generate millions of unique, unrealistic procedurally generated objects, and train a deep neural network to perform grasp planning on these objects. Since the distribution of successful grasps for a given object can be highly multimodal, we propose an autoregressive grasp planning model that maps sensor inputs of a scene to a probability distribution over possible grasps. This model allows us to sample grasps efficiently at test time (or avoid sampling entirely). We evaluate our model architecture and data generation pipeline in simulation and the real world. We find we can achieve a >90% success rate on previously unseen realistic objects at test time in simulation despite having only been trained on random objects. We also demonstrate an 80% success rate on real-world grasp attempts despite having only been trained on random simulated objects.

Dealing with sparse rewards is one of the biggest challenges in Reinforcement Learning (RL). We present a novel technique called Hindsight Experience Replay which allows sample-efficient learning from rewards which are sparse and binary and therefore avoid the need for complicated reward engineering. It can be combined with an arbitrary off-policy RL algorithm and may be seen as a form of implicit curriculum. We demonstrate our approach on the task of manipulating objects with a robotic arm. In particular, we run experiments on three different tasks: pushing, sliding, and pick-and-place, in each case using only binary rewards indicating whether or not the task is completed. Our ablation studies show that Hindsight Experience Replay is a crucial ingredient which makes training possible in these challenging environments. We show that our policies trained on a physics simulation can be deployed on a physical robot and successfully complete the task. The video presenting our experiments is available at https://goo.gl/SMrQnI.

How many labeled examples are needed to estimate a classifier's performance on a new dataset? We study the case where data is plentiful, but labels are expensive. We show that by making a few reasonable assumptions on the structure of the data, it is possible to estimate performance curves, with confidence bounds, using a small number of ground truth labels. Our approach, which we call Semi supervised Performance Evaluation (SPE), is based on a generative model for the classifier's confidence scores. In addition to estimating the performance of classifiers on new datasets, SPE can be used to recalibrate a classifier by re-estimating the class-conditional confidence distributions.

Is it possible to crowdsource categorization? Amongst the challenges: (a) each annotator has only a partial view of the data, (b) different annotators may have different clustering criteria and may produce different numbers of categories, (c) the underlying category structure may be hierarchical. We propose a Bayesian model of how annotators may approach clustering and show how one may infer clusters/categories, as well as annotator parameters, using this model. Our experiments, carried out on large collections of images, suggest that Bayesian crowdclustering works well and may be superior to single-expert annotations.

We present a fast and accurate algorithm for computing the 2D pose of objects in images called cascaded pose regression (CPR). CPR progressively refines a loosely specified initial guess, where each refinement is carried out by a different regressor. Each regressor performs simple image measurements that are dependent on the output of the previous regressors; the entire system is automatically learned from human annotated training examples. CPR is not restricted to rigid transformations: `pose' is any parameterized variation of the object's appearance such as the degrees of freedom of deformable and articulated objects. We compare CPR against both standard regression techniques and human performance (computed from redundant human annotations). Experiments on three diverse datasets (mice, faces, fish) suggest CPR is fast (2-3ms per pose estimate), accurate (approaching human performance), and easy to train from small amounts of labeled data.

Distributing labeling tasks among hundreds or thousands of annotators is an increasingly important method for annotating large datasets. We present a method for estimating the underlying value (e.g. the class) of each image from (noisy) annotations provided by multiple annotators. Our method is based on a model of the image formation and annotation process. Each image has different characteristics that are represented in an abstract Euclidean space. Each annotator is modeled as a multidimensional entity with variables representing competence, expertise and bias. This allows the model to discover and represent groups of annotators that have different sets of skills and knowledge, as well as groups of images that differ qualitatively. We find that our model predicts ground truth labels on both synthetic and real data more accurately than state of the art methods. Experiments also show that our model, starting from a set of binary labels, may discover rich information, such as different "schools of thought" amongst the annotators, and can group together images belonging to separate categories.

Gathering a large collection of images has been made quite easy by social and image sharing websites, e.g. flickr.com. However, using such collections faces the problem that they contain a large number of duplicates and highly similar images. This work tackles the problem of how to automatically organize image collections into sets of similar images, called image families hereinafter. We thoroughly compare the performance of two approaches to measure image similarity: global descriptors vs. a set of local descriptors. We assess the performance of these approaches as the problem scales up to thousands of images and hundreds of families. We present our results on a new dataset of CD/DVD game covers.