The prevalent deployment of learning from human preferences through reinforcement learning (RLHF) relies on two important approximations: the first assumes that pairwise preferences can be substituted with pointwise rewards. The second assumes that a reward model trained on these pointwise rewards can generalize from collected data to out-of-distribution data sampled by the policy. Recently, Direct Preference Optimisation (DPO) has been proposed as an approach that bypasses the second approximation and learn directly a policy from collected data without the reward modelling stage. However, this method still heavily relies on the first approximation. In this paper we try to gain a deeper theoretical understanding of these practical algorithms. In particular we derive a new general objective called $\Psi$PO for learning from human preferences that is expressed in terms of pairwise preferences and therefore bypasses both approximations. This new general objective allows us to perform an in-depth analysis of the behavior of RLHF and DPO (as special cases of $\Psi$PO) and to identify their potential pitfalls. We then consider another special case for $\Psi$PO by setting $\Psi$ simply to Identity, for which we can derive an efficient optimisation procedure, prove performance guarantees and demonstrate its empirical superiority to DPO on some illustrative examples.
相關內容
Continual learning (CL) refers to the ability of an intelligent system to sequentially acquire and retain knowledge from a stream of data with as little computational overhead as possible. To this end; regularization, replay, architecture, and parameter isolation approaches were introduced to the literature. Parameter isolation using a sparse network which enables to allocate distinct parts of the neural network to different tasks and also allows to share of parameters between tasks if they are similar. Dynamic Sparse Training (DST) is a prominent way to find these sparse networks and isolate them for each task. This paper is the first empirical study investigating the effect of different DST components under the CL paradigm to fill a critical research gap and shed light on the optimal configuration of DST for CL if it exists. Therefore, we perform a comprehensive study in which we investigate various DST components to find the best topology per task on well-known CIFAR100 and miniImageNet benchmarks in a task-incremental CL setup since our primary focus is to evaluate the performance of various DST criteria, rather than the process of mask selection. We found that, at a low sparsity level, Erdos-R\'enyi Kernel (ERK) initialization utilizes the backbone more efficiently and allows to effectively learn increments of tasks. At a high sparsity level, unless it is extreme, uniform initialization demonstrates a more reliable and robust performance. In terms of growth strategy; performance is dependent on the defined initialization strategy and the extent of sparsity. Finally, adaptivity within DST components is a promising way for better continual learners.
We propose and demonstrate a compositional framework for training and verifying reinforcement learning (RL) systems within a multifidelity sim-to-real pipeline, in order to deploy reliable and adaptable RL policies on physical hardware. By decomposing complex robotic tasks into component subtasks and defining mathematical interfaces between them, the framework allows for the independent training and testing of the corresponding subtask policies, while simultaneously providing guarantees on the overall behavior that results from their composition. By verifying the performance of these subtask policies using a multifidelity simulation pipeline, the framework not only allows for efficient RL training, but also for a refinement of the subtasks and their interfaces in response to challenges arising from discrepancies between simulation and reality. In an experimental case study we apply the framework to train and deploy a compositional RL system that successfully pilots a Warthog unmanned ground robot.
Federated learning (FL) allows multiple clients to collaboratively learn a globally shared model through cycles of model aggregation and local model training, without the need to share data. Most existing FL methods train local models separately on different clients, and then simply average their parameters to obtain a centralized model on the server side. However, these approaches generally suffer from large aggregation errors and severe local forgetting, which are particularly bad in heterogeneous data settings. To tackle these issues, in this paper, we propose a novel FL framework that uses online Laplace approximation to approximate posteriors on both the client and server side. On the server side, a multivariate Gaussian product mechanism is employed to construct and maximize a global posterior, largely reducing the aggregation errors induced by large discrepancies between local models. On the client side, a prior loss that uses the global posterior probabilistic parameters delivered from the server is designed to guide the local training. Binding such learning constraints from other clients enables our method to mitigate local forgetting. Finally, we achieve state-of-the-art results on several benchmarks, clearly demonstrating the advantages of the proposed method.
Online Continual Learning (CL) solves the problem of learning the ever-emerging new classification tasks from a continuous data stream. Unlike its offline counterpart, in online CL, the training data can only be seen once. Most existing online CL research regards catastrophic forgetting (i.e., model stability) as almost the only challenge. In this paper, we argue that the model's capability to acquire new knowledge (i.e., model plasticity) is another challenge in online CL. While replay-based strategies have been shown to be effective in alleviating catastrophic forgetting, there is a notable gap in research attention toward improving model plasticity. To this end, we propose Collaborative Continual Learning (CCL), a collaborative learning based strategy to improve the model's capability in acquiring new concepts. Additionally, we introduce Distillation Chain (DC), a novel collaborative learning scheme to boost the training of the models. We adapted CCL-DC to existing representative online CL works. Extensive experiments demonstrate that even if the learners are well-trained with state-of-the-art online CL methods, our strategy can still improve model plasticity dramatically, and thereby improve the overall performance by a large margin.
A major concern in using deep learning based generative models for document-grounded dialogs is the potential generation of responses that are not \textit{faithful} to the underlying document. Existing automated metrics used for evaluating the faithfulness of response with respect to the grounding document measure the degree of similarity between the generated response and the document's content. However, these automated metrics are far from being well aligned with human judgments. Therefore, to improve the measurement of faithfulness, we propose a new metric that utilizes (Conditional) Point-wise Mutual Information (PMI) between the generated response and the source document, conditioned on the dialogue. PMI quantifies the extent to which the document influences the generated response -- with a higher PMI indicating a more faithful response. We build upon this idea to create a new decoding technique that incorporates PMI into the response generation process to predict more faithful responses. Our experiments on the BEGIN benchmark demonstrate an improved correlation of our metric with human evaluation. We also show that our decoding technique is effective in generating more faithful responses when compared to standard decoding techniques on a set of publicly available document-grounded dialog datasets.
Preference-based feedback is important for many applications in reinforcement learning where direct evaluation of a reward function is not feasible. A notable recent example arises in reinforcement learning from human feedback (RLHF) on large language models. For many applications of RLHF, the cost of acquiring the human feedback can be substantial. In this work, we take advantage of the fact that one can often choose contexts at which to obtain human feedback in order to most efficiently identify a good policy, and formalize this as an offline contextual dueling bandit problem. We give an upper-confidence-bound style algorithm for this problem and prove a polynomial worst-case regret bound. We then provide empirical confirmation in a synthetic setting that our approach outperforms existing methods. After, we extend the setting and methodology for practical use in RLHF training of large language models. Here, our method is able to reach better performance with fewer samples of human preferences than multiple baselines on three real-world datasets.
The foundation models (FMs) have been used to generate synthetic public datasets for the heterogeneous federated learning (HFL) problem where each client uses a unique model architecture. However, the vulnerabilities of integrating FMs, especially against backdoor attacks, are not well-explored in the HFL contexts. In this paper, we introduce a novel backdoor attack mechanism for HFL that circumvents the need for client compromise or ongoing participation in the FL process. This method plants and transfers the backdoor through a generated synthetic public dataset, which could help evade existing backdoor defenses in FL by presenting normal client behaviors. Empirical experiments across different HFL configurations and benchmark datasets demonstrate the effectiveness of our attack compared to traditional client-based attacks. Our findings reveal significant security risks in developing robust FM-assisted HFL systems. This research contributes to enhancing the safety and integrity of FL systems, highlighting the need for advanced security measures in the era of FMs.
Recent contrastive representation learning methods rely on estimating mutual information (MI) between multiple views of an underlying context. E.g., we can derive multiple views of a given image by applying data augmentation, or we can split a sequence into views comprising the past and future of some step in the sequence. Contrastive lower bounds on MI are easy to optimize, but have a strong underestimation bias when estimating large amounts of MI. We propose decomposing the full MI estimation problem into a sum of smaller estimation problems by splitting one of the views into progressively more informed subviews and by applying the chain rule on MI between the decomposed views. This expression contains a sum of unconditional and conditional MI terms, each measuring modest chunks of the total MI, which facilitates approximation via contrastive bounds. To maximize the sum, we formulate a contrastive lower bound on the conditional MI which can be approximated efficiently. We refer to our general approach as Decomposed Estimation of Mutual Information (DEMI). We show that DEMI can capture a larger amount of MI than standard non-decomposed contrastive bounds in a synthetic setting, and learns better representations in a vision domain and for dialogue generation.
Few sample learning (FSL) is significant and challenging in the field of machine learning. The capability of learning and generalizing from very few samples successfully is a noticeable demarcation separating artificial intelligence and human intelligence since humans can readily establish their cognition to novelty from just a single or a handful of examples whereas machine learning algorithms typically entail hundreds or thousands of supervised samples to guarantee generalization ability. Despite the long history dated back to the early 2000s and the widespread attention in recent years with booming deep learning technologies, little surveys or reviews for FSL are available until now. In this context, we extensively review 200+ papers of FSL spanning from the 2000s to 2019 and provide a timely and comprehensive survey for FSL. In this survey, we review the evolution history as well as the current progress on FSL, categorize FSL approaches into the generative model based and discriminative model based kinds in principle, and emphasize particularly on the meta learning based FSL approaches. We also summarize several recently emerging extensional topics of FSL and review the latest advances on these topics. Furthermore, we highlight the important FSL applications covering many research hotspots in computer vision, natural language processing, audio and speech, reinforcement learning and robotic, data analysis, etc. Finally, we conclude the survey with a discussion on promising trends in the hope of providing guidance and insights to follow-up researches.
Deep learning has yielded state-of-the-art performance on many natural language processing tasks including named entity recognition (NER). However, this typically requires large amounts of labeled data. In this work, we demonstrate that the amount of labeled training data can be drastically reduced when deep learning is combined with active learning. While active learning is sample-efficient, it can be computationally expensive since it requires iterative retraining. To speed this up, we introduce a lightweight architecture for NER, viz., the CNN-CNN-LSTM model consisting of convolutional character and word encoders and a long short term memory (LSTM) tag decoder. The model achieves nearly state-of-the-art performance on standard datasets for the task while being computationally much more efficient than best performing models. We carry out incremental active learning, during the training process, and are able to nearly match state-of-the-art performance with just 25\% of the original training data.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191