将来，强大的AI系统可能会在高风险的设置中部署，在这种情况下，单个故障可能是灾难性的。在高风险设置中改善AI安全性的一种技术是对手训练，该培训使用对手来生成示例进行训练，以实现更好的最差表现。在这项工作中，我们将语言生成任务用作测试台，以通过对抗性培训来实现高可靠性。我们创建了一系列的对抗训练技术 - 包括一种有助于人类对手的工具 - 以在分类器中找到和消除故障，该分类器过滤了发电机建议的文本完成。在简单的“避免受伤”任务中，我们确定我们可以设置非常保守的分类器阈值，而不会显着影响过滤后的输出的质量。使用我们选择的阈值，使用基线分类器进行过滤，将不安全完成的速度从分布数据的数据降低到约2.4％至0.003％，这是我们测量能力的极限。我们发现，对抗性训练可显着提高对我们训练的对抗攻击的鲁棒性，而不会影响分布性能。我们希望在高风险的可靠性环境中看到进一步的工作，包括更强大的工具来增强人类对手，以及更好的方法来衡量高水平的可靠性，直到我们可以自信地排除强大模型的灾难性部署时间失败的可能性。

Extracting complex structures from grid-based data is a common key step in automated medical image analysis. The conventional solution to recovering tree-structured geometries typically involves computing the minimal cost path through intermediate representations derived from segmentation masks. However, this methodology has significant limitations in the context of projective imaging of tree-structured 3D anatomical data such as coronary arteries, since there are often overlapping branches in the 2D projection. In this work, we propose a novel approach to predicting tree connectivity structure which reformulates the task as an optimization problem over individual steps of a recursive process. We design and train a two-stage model which leverages the UNet and Transformer architectures and introduces an image-based prompting technique. Our proposed method achieves compelling results on a pair of synthetic datasets, and outperforms a shortest-path baseline.

Curriculum learning and self-paced learning are the training strategies that gradually feed the samples from easy to more complex. They have captivated increasing attention due to their excellent performance in robotic vision. Most recent works focus on designing curricula based on difficulty levels in input samples or smoothing the feature maps. However, smoothing labels to control the learning utility in a curriculum manner is still unexplored. In this work, we design a paced curriculum by label smoothing (P-CBLS) using paced learning with uniform label smoothing (ULS) for classification tasks and fuse uniform and spatially varying label smoothing (SVLS) for semantic segmentation tasks in a curriculum manner. In ULS and SVLS, a bigger smoothing factor value enforces a heavy smoothing penalty in the true label and limits learning less information. Therefore, we design the curriculum by label smoothing (CBLS). We set a bigger smoothing value at the beginning of training and gradually decreased it to zero to control the model learning utility from lower to higher. We also designed a confidence-aware pacing function and combined it with our CBLS to investigate the benefits of various curricula. The proposed techniques are validated on four robotic surgery datasets of multi-class, multi-label classification, captioning, and segmentation tasks. We also investigate the robustness of our method by corrupting validation data into different severity levels. Our extensive analysis shows that the proposed method improves prediction accuracy and robustness.

Temporal reasoning is the task of predicting temporal relations of event pairs with corresponding contexts. While some temporal reasoning models perform reasonably well on in-domain benchmarks, we have little idea of the systems' generalizability due to existing datasets' limitations. In this work, we introduce a novel task named TODAY that bridges this gap with temporal differential analysis, which as the name suggests, evaluates if systems can correctly understand the effect of incremental changes. Specifically, TODAY makes slight context changes for given event pairs, and systems need to tell how this subtle contextual change will affect temporal relation distributions. To facilitate learning, TODAY also annotates human explanations. We show that existing models, including GPT-3, drop to random guessing on TODAY, suggesting that they heavily rely on spurious information rather than proper reasoning for temporal predictions. On the other hand, we show that TODAY's supervision style and explanation annotations can be used in joint learning and encourage models to use more appropriate signals during training and outperform across several benchmarks. TODAY can also be used to train models to solicit incidental supervision from noisy sources such as GPT-3 and moves farther towards generic temporal reasoning systems.

State-of-the-art 3D semantic segmentation models are trained on the off-the-shelf public benchmarks, but they often face the major challenge when these well-trained models are deployed to a new domain. In this paper, we propose an Active-and-Adaptive Segmentation (ADAS) baseline to enhance the weak cross-domain generalization ability of a well-trained 3D segmentation model, and bridge the point distribution gap between domains. Specifically, before the cross-domain adaptation stage begins, ADAS performs an active sampling operation to select a maximally-informative subset from both source and target domains for effective adaptation, reducing the adaptation difficulty under 3D scenarios. Benefiting from the rise of multi-modal 2D-3D datasets, ADAS utilizes a cross-modal attention-based feature fusion module that can extract a representative pair of image features and point features to achieve a bi-directional image-point feature interaction for better safe adaptation. Experimentally, ADAS is verified to be effective in many cross-domain settings including: 1) Unsupervised Domain Adaptation (UDA), which means that all samples from target domain are unlabeled; 2) Unsupervised Few-shot Domain Adaptation (UFDA) which means that only a few unlabeled samples are available in the unlabeled target domain; 3) Active Domain Adaptation (ADA) which means that the selected target samples by ADAS are manually annotated. Their results demonstrate that ADAS achieves a significant accuracy gain by easily coupling ADAS with self-training methods or off-the-shelf UDA works.

As language models (LMs) scale, they develop many novel behaviors, good and bad, exacerbating the need to evaluate how they behave. Prior work creates evaluations with crowdwork (which is time-consuming and expensive) or existing data sources (which are not always available). Here, we automatically generate evaluations with LMs. We explore approaches with varying amounts of human effort, from instructing LMs to write yes/no questions to making complex Winogender schemas with multiple stages of LM-based generation and filtering. Crowdworkers rate the examples as highly relevant and agree with 90-100% of labels, sometimes more so than corresponding human-written datasets. We generate 154 datasets and discover new cases of inverse scaling where LMs get worse with size. Larger LMs repeat back a dialog user's preferred answer ("sycophancy") and express greater desire to pursue concerning goals like resource acquisition and goal preservation. We also find some of the first examples of inverse scaling in RL from Human Feedback (RLHF), where more RLHF makes LMs worse. For example, RLHF makes LMs express stronger political views (on gun rights and immigration) and a greater desire to avoid shut down. Overall, LM-written evaluations are high-quality and let us quickly discover many novel LM behaviors.

In this paper, we discuss an imitation learning based method for reducing the calibration error for a mixed reality system consisting of a vision sensor and a projector. Unlike a head mounted display, in this setup, augmented information is available to a human subject via the projection of a scene into the real world. Inherently, the camera and projector need to be calibrated as a stereo setup to project accurate information in 3D space. Previous calibration processes require multiple recording and parameter tuning steps to achieve the desired calibration, which is usually time consuming process. In order to avoid such tedious calibration, we train a CNN model to iteratively correct the extrinsic offset given a QR code and a projected pattern. We discuss the overall system setup, data collection for training, and results of the auto-correction model.

Background samples provide key contextual information for segmenting regions of interest (ROIs). However, they always cover a diverse set of structures, causing difficulties for the segmentation model to learn good decision boundaries with high sensitivity and precision. The issue concerns the highly heterogeneous nature of the background class, resulting in multi-modal distributions. Empirically, we find that neural networks trained with heterogeneous background struggle to map the corresponding contextual samples to compact clusters in feature space. As a result, the distribution over background logit activations may shift across the decision boundary, leading to systematic over-segmentation across different datasets and tasks. In this study, we propose context label learning (CoLab) to improve the context representations by decomposing the background class into several subclasses. Specifically, we train an auxiliary network as a task generator, along with the primary segmentation model, to automatically generate context labels that positively affect the ROI segmentation accuracy. Extensive experiments are conducted on several challenging segmentation tasks and datasets. The results demonstrate that CoLab can guide the segmentation model to map the logits of background samples away from the decision boundary, resulting in significantly improved segmentation accuracy. Code is available.

As AI systems become more capable, we would like to enlist their help to supervise other AIs. We experiment with methods for training a harmless AI assistant through self-improvement, without any human labels identifying harmful outputs. The only human oversight is provided through a list of rules or principles, and so we refer to the method as 'Constitutional AI'. The process involves both a supervised learning and a reinforcement learning phase. In the supervised phase we sample from an initial model, then generate self-critiques and revisions, and then finetune the original model on revised responses. In the RL phase, we sample from the finetuned model, use a model to evaluate which of the two samples is better, and then train a preference model from this dataset of AI preferences. We then train with RL using the preference model as the reward signal, i.e. we use 'RL from AI Feedback' (RLAIF). As a result we are able to train a harmless but non-evasive AI assistant that engages with harmful queries by explaining its objections to them. Both the SL and RL methods can leverage chain-of-thought style reasoning to improve the human-judged performance and transparency of AI decision making. These methods make it possible to control AI behavior more precisely and with far fewer human labels.

We present a generalised architecture for reactive mobile manipulation while a robot's base is in motion toward the next objective in a high-level task. By performing tasks on-the-move, overall cycle time is reduced compared to methods where the base pauses during manipulation. Reactive control of the manipulator enables grasping objects with unpredictable motion while improving robustness against perception errors, environmental disturbances, and inaccurate robot control compared to open-loop, trajectory-based planning approaches. We present an example implementation of the architecture and investigate the performance on a series of pick and place tasks with both static and dynamic objects and compare the performance to baseline methods. Our method demonstrated a real-world success rate of over 99%, failing in only a single trial from 120 attempts with a physical robot system. The architecture is further demonstrated on other mobile manipulator platforms in simulation. Our approach reduces task time by up to 48%, while also improving reliability, gracefulness, and predictability compared to existing architectures for mobile manipulation. See https://benburgesslimerick.github.io/ManipulationOnTheMove for supplementary materials.