Yu Sun

Renhao Wang, Yu Sun, Arnuv Tandon et al. · berkeley, ibm-research

Prior work has established Test-Time Training (TTT) as a general framework to further improve a trained model at test time. Before making a prediction on each test instance, the model is first trained on the same instance using a self-supervised task such as reconstruction. We extend TTT to the streaming setting, where multiple test instances - video frames in our case - arrive in temporal order. Our extension is online TTT: The current model is initialized from the previous model, then trained on the current frame and a small window of frames immediately before. Online TTT significantly outperforms the fixed-model baseline for four tasks, on three real-world datasets. The improvements are more than 2.2x and 1.5x for instance and panoptic segmentation. Surprisingly, online TTT also outperforms its offline variant that accesses strictly more information, training on all frames from the entire test video regardless of temporal order. This finding challenges those in prior work using synthetic videos. We formalize a notion of locality as the advantage of online over offline TTT, and analyze its role with ablations and a theory based on bias-variance trade-off.

Md Sadman Sakib, Yu Sun

Task planning for robotic cooking involves generating a sequence of actions for a robot to prepare a meal successfully. This paper introduces a novel task tree generation pipeline producing correct planning and efficient execution for cooking tasks. Our method first uses a large language model (LLM) to retrieve recipe instructions and then utilizes a fine-tuned GPT-3 to convert them into a task tree, capturing sequential and parallel dependencies among subtasks. The pipeline then mitigates the uncertainty and unreliable features of LLM outputs using task tree retrieval. We combine multiple LLM task tree outputs into a graph and perform a task tree retrieval to avoid questionable nodes and high-cost nodes to improve planning correctness and improve execution efficiency. Our evaluation results show its superior performance compared to previous works in task planning accuracy and efficiency.

Md Sadman Sakib, Yu Sun

The inherent probabilistic nature of Large Language Models (LLMs) introduces an element of unpredictability, raising concerns about potential discrepancies in their output. This paper introduces an innovative approach aims to generate correct and optimal robotic task plans for diverse real-world demands and scenarios. LLMs have been used to generate task plans, but they are unreliable and may contain wrong, questionable, or high-cost steps. The proposed approach uses LLM to generate a number of task plans as trees and amalgamates them into a graph by removing questionable paths. Then an optimal task tree can be retrieved to circumvent questionable and high-cost nodes, thereby improving planning accuracy and execution efficiency. The approach is further improved by incorporating a large knowledge network. Leveraging GPT-4 further, the high-level task plan is converted into a low-level Planning Domain Definition Language (PDDL) plan executable by a robot. Evaluation results highlight the superior accuracy and efficiency of our approach compared to previous methodologies in the field of task planning.

Yu Sun, Gaojian Xiong, Xianxun Yao et al.

Deep gradient inversion attacks expose a serious threat to Federated Learning (FL) by accurately recovering private data from shared gradients. However, the state-of-the-art heavily relies on impractical assumptions to access excessive auxiliary data, which violates the basic data partitioning principle of FL. In this paper, a novel method, Gradient Inversion Attack using Practical Image Prior (GI-PIP), is proposed under a revised threat model. GI-PIP exploits anomaly detection models to capture the underlying distribution from fewer data, while GAN-based methods consume significant more data to synthesize images. The extracted distribution is then leveraged to regulate the attack process as Anomaly Score loss. Experimental results show that GI-PIP achieves a 16.12 dB PSNR recovery using only 3.8% data of ImageNet, while GAN-based methods necessitate over 70%. Moreover, GI-PIP exhibits superior capability on distribution generalization compared to GAN-based methods. Our approach significantly alleviates the auxiliary data requirement on both amount and distribution in gradient inversion attacks, hence posing more substantial threat to real-world FL.

Jun Wu, Xuesong Ye

The proliferation of fake news poses a serious threat to society, as it can misinform and manipulate the public, erode trust in institutions, and undermine democratic processes. To address this issue, we present FakeSwarm, a fake news identification system that leverages the swarming characteristics of fake news. To extract the swarm behavior, we propose a novel concept of fake news swarming characteristics and design three types of swarm features, including principal component analysis, metric representation, and position encoding. We evaluate our system on a public dataset and demonstrate the effectiveness of incorporating swarm features in fake news identification, achieving an f1-score and accuracy of over 97% by combining all three types of swarm features. Furthermore, we design an online learning pipeline based on the hypothesis of the temporal distribution pattern of fake news emergence, validated on a topic with early emerging fake news and a shortage of text samples, showing that swarm features can significantly improve recall rates in such cases. Our work provides a new perspective and approach to fake news detection and highlights the importance of considering swarming characteristics in detecting fake news.

Moritz Hardt, Yu Sun

Many recent efforts augment language models with retrieval, by adding retrieved data to the input context. For this approach to succeed, the retrieved data must be added at both training and test time. Moreover, as input length grows linearly with the size of retrieved data, cost in computation and memory grows quadratically for modern Transformers. To avoid these complications, we simply fine-tune the model on retrieved data at test time, using its standard training setup. We build a large-scale distributed index based on text embeddings of the Pile dataset. For each test input, our system retrieves its neighbors and fine-tunes the model on their text. Surprisingly, retrieving and training on as few as 20 neighbors, each for only one gradient iteration, drastically improves performance across more than 20 language modeling tasks in the Pile. For example, test-time training with nearest neighbors significantly narrows the performance gap between a small GPT-2 and a GPT-Neo model more than 10 times larger. Sufficient index quality and size, however, are necessary. Our work establishes a first baseline of test-time training for language modeling.

Nicklas Hansen, Rishabh Jangir, Yu Sun et al.

In most real world scenarios, a policy trained by reinforcement learning in one environment needs to be deployed in another, potentially quite different environment. However, generalization across different environments is known to be hard. A natural solution would be to keep training after deployment in the new environment, but this cannot be done if the new environment offers no reward signal. Our work explores the use of self-supervision to allow the policy to continue training after deployment without using any rewards. While previous methods explicitly anticipate changes in the new environment, we assume no prior knowledge of those changes yet still obtain significant improvements. Empirical evaluations are performed on diverse simulation environments from DeepMind Control suite and ViZDoom, as well as real robotic manipulation tasks in continuously changing environments, taking observations from an uncalibrated camera. Our method improves generalization in 31 out of 36 environments across various tasks and outperforms domain randomization on a majority of environments.

Yu Sun, Eric Tzeng, Trevor Darrell et al.

This paper addresses unsupervised domain adaptation, the setting where labeled training data is available on a source domain, but the goal is to have good performance on a target domain with only unlabeled data. Like much of previous work, we seek to align the learned representations of the source and target domains while preserving discriminability. The way we accomplish alignment is by learning to perform auxiliary self-supervised task(s) on both domains simultaneously. Each self-supervised task brings the two domains closer together along the direction relevant to that task. Training this jointly with the main task classifier on the source domain is shown to successfully generalize to the unlabeled target domain. The presented objective is straightforward to implement and easy to optimize. We achieve state-of-the-art results on four out of seven standard benchmarks, and competitive results on segmentation adaptation. We also demonstrate that our method composes well with another popular pixel-level adaptation method.

Ahmad Babaeian Jelodar, Md Sirajus Salekin, Yu Sun

Understanding object states is as important as object recognition for robotic task planning and manipulation. To our knowledge, this paper explicitly introduces and addresses the state identification problem in cooking related images for the first time. In this paper, objects and ingredients in cooking videos are explored and the most frequent objects are analyzed. Eleven states from the most frequent cooking objects are examined and a dataset of images containing those objects and their states is created. As a solution to the state identification problem, a Resnet based deep model is proposed. The model is initialized with Imagenet weights and trained on the dataset of eleven classes. The trained state identification model is evaluated on a subset of the Imagenet dataset and state labels are provided using a combination of the model with manual checking. Moreover, an individual model is fine-tuned for each object in the dataset using the weights from the initially trained model and object-specific images, where significant improvement is demonstrated.

Chuan Guo, Geoff Pleiss, Yu Sun et al.

Confidence calibration -- the problem of predicting probability estimates representative of the true correctness likelihood -- is important for classification models in many applications. We discover that modern neural networks, unlike those from a decade ago, are poorly calibrated. Through extensive experiments, we observe that depth, width, weight decay, and Batch Normalization are important factors influencing calibration. We evaluate the performance of various post-processing calibration methods on state-of-the-art architectures with image and document classification datasets. Our analysis and experiments not only offer insights into neural network learning, but also provide a simple and straightforward recipe for practical settings: on most datasets, temperature scaling -- a single-parameter variant of Platt Scaling -- is surprisingly effective at calibrating predictions.

Xilun Chen, Yu Sun, Ben Athiwaratkun et al.

In recent years great success has been achieved in sentiment classification for English, thanks in part to the availability of copious annotated resources. Unfortunately, most languages do not enjoy such an abundance of labeled data. To tackle the sentiment classification problem in low-resource languages without adequate annotated data, we propose an Adversarial Deep Averaging Network (ADAN) to transfer the knowledge learned from labeled data on a resource-rich source language to low-resource languages where only unlabeled data exists. ADAN has two discriminative branches: a sentiment classifier and an adversarial language discriminator. Both branches take input from a shared feature extractor to learn hidden representations that are simultaneously indicative for the classification task and invariant across languages. Experiments on Chinese and Arabic sentiment classification demonstrate that ADAN significantly outperforms state-of-the-art systems.

Gao Huang, Yu Sun, Zhuang Liu et al.

Very deep convolutional networks with hundreds of layers have led to significant reductions in error on competitive benchmarks. Although the unmatched expressiveness of the many layers can be highly desirable at test time, training very deep networks comes with its own set of challenges. The gradients can vanish, the forward flow often diminishes, and the training time can be painfully slow. To address these problems, we propose stochastic depth, a training procedure that enables the seemingly contradictory setup to train short networks and use deep networks at test time. We start with very deep networks but during training, for each mini-batch, randomly drop a subset of layers and bypass them with the identity function. This simple approach complements the recent success of residual networks. It reduces training time substantially and improves the test error significantly on almost all data sets that we used for evaluation. With stochastic depth we can increase the depth of residual networks even beyond 1200 layers and still yield meaningful improvements in test error (4.91% on CIFAR-10).