Existing NAS methods suffer from either an excessive amount of time for repetitive sampling and training of many task-irrelevant architectures. To tackle such limitations of existing NAS methods, we propose a paradigm shift from NAS to a novel conditional Neural Architecture Generation (NAG) framework based on diffusion models, dubbed DiffusionNAG. Specifically, we consider the neural architectures as directed graphs and propose a graph diffusion model for generating them. Moreover, with the guidance of parameterized predictors, DiffusionNAG can flexibly generate task-optimal architectures with the desired properties for diverse tasks, by sampling from a region that is more likely to satisfy the properties. This conditional NAG scheme is significantly more efficient than previous NAS schemes which sample the architectures and filter them using the property predictors. We validate the effectiveness of DiffusionNAG through extensive experiments in two predictor-based NAS scenarios: Transferable NAS and Bayesian Optimization (BO)-based NAS. DiffusionNAG achieves superior performance with speedups of up to 35 times when compared to the baselines on Transferable NAS benchmarks. Furthermore, when integrated into a BO-based algorithm, DiffusionNAG outperforms existing BO-based NAS approaches, particularly in the large MobileNetV3 search space on the ImageNet 1K dataset. Code is available at https://github.com/CownowAn/DiffusionNAG.
There have been wide spread claims in the literature about the emergent reasoning capabilities of Pretrained Large Language Models. However, recent studies, have found that their ability to plan remains questionable. Through our experiments using GPT-2, we empirically demonstrate that the performance of a finetuned baseline remains poor because it violates pre-conditions of actions in the plans that it generates. To improve the planning capabilities of a finetuned LLM, we train a verifier, which can classify actions as being valid or invalid in a particular state. By randomly sampling actions from the same dataset, we generate examples of invalid actions which are then used to train a verifier which can check for action applicability. In the presence of diverse sampling from a generator and a verifier which can prune invalid trajectories, we show significant gains in the success rate on the Blocksworld domain. Additionally, we show that finetuning the GPT-2 generator itself to create the verifier generalizes better than finetuning the base GPT-2. Lastly, we investigate the role of the sampling temperature which can be used to control the exploration-exploitation tradeoff.
Question generation (QG) is the task of generating a valid and fluent question based on a given context and the target answer. According to various purposes, even given the same context, instructors can ask questions about different concepts, and even the same concept can be written in different ways. However, the evaluation for QG usually depends on single reference-based similarity metrics, such as n-gram-based metric or learned metric, which is not sufficient to fully evaluate the potential of QG methods. To this end, we propose to paraphrase the reference question for a more robust QG evaluation. Using large language models such as GPT-3, we created semantically and syntactically diverse questions, then adopt the simple aggregation of the popular evaluation metrics as the final scores. Through our experiments, we found that using multiple (pseudo) references is more effective for QG evaluation while showing a higher correlation with human evaluations than evaluation with a single reference.
Meta-Bayesian optimisation (meta-BO) aims to improve the sample efficiency of Bayesian optimisation by leveraging data from related tasks. While previous methods successfully meta-learn either a surrogate model or an acquisition function independently, joint training of both components remains an open challenge. This paper proposes the first end-to-end differentiable meta-BO framework that generalises neural processes to learn acquisition functions via transformer architectures. We enable this end-to-end framework with reinforcement learning (RL) to tackle the lack of labelled acquisition data. Early on, we notice that training transformer-based neural processes from scratch with RL is challenging due to insufficient supervision, especially when rewards are sparse. We formalise this claim with a combinatorial analysis showing that the widely used notion of regret as a reward signal exhibits a logarithmic sparsity pattern in trajectory lengths. To tackle this problem, we augment the RL objective with an auxiliary task that guides part of the architecture to learn a valid probabilistic model as an inductive bias. We demonstrate that our method achieves state-of-the-art regret results against various baselines in experiments on standard hyperparameter optimisation tasks and also outperforms others in the real-world problems of mixed-integer programming tuning, antibody design, and logic synthesis for electronic design automation.
Large language models (LLMs) are excellent in-context learners. However, the sensitivity of data contained in prompts raises privacy concerns. Our work first shows that these concerns are valid: we instantiate a simple but highly effective membership inference attack against the data used to prompt LLMs. To address this vulnerability, one could forego prompting and resort to fine-tuning LLMs with known algorithms for private gradient descent. However, this comes at the expense of the practicality and efficiency offered by prompting. Therefore, we propose to privately learn to prompt. We first show that soft prompts can be obtained privately through gradient descent on downstream data. However, this is not the case for discrete prompts. Thus, we orchestrate a noisy vote among an ensemble of LLMs presented with different prompts, i.e., a flock of stochastic parrots. The vote privately transfers the flock's knowledge into a single public prompt. We show that LLMs prompted with our private algorithms closely match the non-private baselines. For example, using GPT3 as the base model, we achieve a downstream accuracy of 92.7% on the sst2 dataset with ($\epsilon=0.147, \delta=10^{-6}$)-differential privacy vs. 95.2% for the non-private baseline. Through our experiments, we also show that our prompt-based approach is easily deployed with existing commercial APIs.
Segment anything model (SAM) has presented impressive objectness identification capability with the idea of prompt learning and a new collected large-scale dataset. Given a prompt (e.g., points, bounding boxes, or masks) and an input image, SAM is able to generate valid segment masks for all objects indicated by the prompts, presenting high generalization across diverse scenarios and being a general method for zero-shot transfer to downstream vision tasks. Nevertheless, it remains unclear whether SAM may introduce errors in certain threatening scenarios. Clarifying this is of significant importance for applications that require robustness, such as autonomous vehicles. In this paper, we aim to study the testing-time robustness of SAM under adversarial scenarios and common corruptions. To this end, we first build a testing-time robustness evaluation benchmark for SAM by integrating existing public datasets. Second, we extend representative adversarial attacks against SAM and study the influence of different prompts on robustness. Third, we study the robustness of SAM under diverse corruption types by evaluating SAM on corrupted datasets with different prompts. With experiments conducted on SA-1B and KITTI datasets, we find that SAM exhibits remarkable robustness against various corruptions, except for blur-related corruption. Furthermore, SAM remains susceptible to adversarial attacks, particularly when subjected to PGD and BIM attacks. We think such a comprehensive study could highlight the importance of the robustness issues of SAM and trigger a series of new tasks for SAM as well as downstream vision tasks.
Off-policy learning (OPL) aims at finding improved policies from logged bandit data, often by minimizing the inverse propensity scoring (IPS) estimator of the risk. In this work, we investigate a smooth regularization for IPS, for which we derive a two-sided PAC-Bayes generalization bound. The bound is tractable, scalable, interpretable and provides learning certificates. In particular, it is also valid for standard IPS without making the assumption that the importance weights are bounded. We demonstrate the relevance of our approach and its favorable performance through a set of learning tasks. Since our bound holds for standard IPS, we are able to provide insight into when regularizing IPS is useful. Namely, we identify cases where regularization might not be needed. This goes against the belief that, in practice, clipped IPS often enjoys favorable performance than standard IPS in OPL.
Annotating data via crowdsourcing is time-consuming and expensive. Due to these costs, dataset creators often have each annotator label only a small subset of the data. This leads to sparse datasets with examples that are marked by few annotators. The downside of this process is that if an annotator doesn't get to label a particular example, their perspective on it is missed. This is especially concerning for subjective NLP datasets where there is no single correct label: people may have different valid opinions. Thus, we propose using imputation methods to generate the opinions of all annotators for all examples, creating a dataset that does not leave out any annotator's view. We then train and prompt models, using data from the imputed dataset, to make predictions about the distribution of responses and individual annotations. In our analysis of the results, we found that the choice of imputation method significantly impacts soft label changes and distribution. While the imputation introduces noise in the prediction of the original dataset, it has shown potential in enhancing shots for prompts, particularly for low-response-rate annotators. We have made all of our code and data publicly available.
Natural language understanding (NLU) studies often exaggerate or underestimate the capabilities of systems, thereby limiting the reproducibility of their findings. These erroneous evaluations can be attributed to the difficulty of defining and testing NLU adequately. In this position paper, we reconsider this challenge by identifying two types of researcher degrees of freedom. We revisit Turing's original interpretation of the Turing test and indicate that an NLU test does not provide an operational definition; it merely provides inductive evidence that the test subject understands the language sufficiently well to meet stakeholder objectives. In other words, stakeholders are free to arbitrarily define NLU through their objectives. To use the test results as inductive evidence, stakeholders must carefully assess if the interpretation of test scores is valid or not. However, designing and using NLU tests involve other degrees of freedom, such as specifying target skills and defining evaluation metrics. As a result, achieving consensus among stakeholders becomes difficult. To resolve this issue, we propose a validity argument, which is a framework comprising a series of validation criteria across test components. By demonstrating that current practices in NLU studies can be associated with those criteria and organizing them into a comprehensive checklist, we prove that the validity argument can serve as a coherent guideline for designing credible test sets and facilitating scientific communication.
We explore whether Large Language Models (LLMs) are capable of logical reasoning with distorted facts, which we call Deduction under Perturbed Evidence (DUPE). DUPE presents a unique challenge to LLMs since they typically rely on their parameters, which encode mostly accurate information, to reason and make inferences. However, in DUPE, LLMs must reason over manipulated or falsified evidence present in their prompts, which can result in false conclusions that are valid only under the manipulated evidence. Our goal with DUPE is to determine whether LLMs can arrive at these false conclusions and identify whether the dominant factor influencing the deduction process is the encoded data in the parameters or the manipulated evidence in the prompts. To evaluate the DUPE capabilities of LLMs, we create a DUPEd version of the StrategyQA dataset, where facts are manipulated to reverse the answer to the question. Our findings show that even the most advanced GPT models struggle to reason on manipulated facts - showcasing poor DUPE skills - with accuracy dropping by 45% compared to the original dataset. We also investigate prompt settings inspired from student simulation models, which mitigate the accuracy drop to some extent. Our findings have practical implications for understanding the performance of LLMs in real-world applications such as student simulation models that involve reasoning over inaccurate information.
We address a fundamental challenge in Natural Language Generation (NLG) model evaluation -- the design and evaluation of evaluation metrics. Recognizing the limitations of existing automatic metrics and noises from how current human evaluation was conducted, we propose MetricEval, a framework informed by measurement theory, the foundation of educational test design, for conceptualizing and evaluating the reliability and validity of NLG evaluation metrics. The framework formalizes the source of measurement error and offers statistical tools for evaluating evaluation metrics based on empirical data. With our framework, one can quantify the uncertainty of the metrics to better interpret the result. To exemplify the use of our framework in practice, we analyzed a set of evaluation metrics for summarization and identified issues related to conflated validity structure in human-eval and reliability in LLM-based metrics. Through MetricEval, we aim to promote the design, evaluation, and interpretation of valid and reliable metrics to advance robust and effective NLG models.
Manifolds discovered by machine learning models provide a compact representation of the underlying data. Geodesics on these manifolds define locally length-minimising curves and provide a notion of distance, which are key for reduced-order modelling, statistical inference, and interpolation. In this work, we first analyse existing methods for computing length-minimising geodesics. We find that these are not suitable for obtaining valid paths, and thus, geodesic distances. We remedy these shortcomings by leveraging numerical tools from differential geometry, which provide the means to obtain Hamiltonian-conserving geodesics. Second, we propose a model-based parameterisation for distance fields and geodesic flows on continuous manifolds. Our approach exploits a manifold-aware extension to the Eikonal equation, eliminating the need for approximations or discretisation. Finally, we develop a curvature-based training mechanism, sampling and scaling points in regions of the manifold exhibiting larger values of the Ricci scalar. This sampling and scaling approach ensures that we capture regions of the manifold subject to higher degrees of geodesic deviation. Our proposed methods provide principled means to compute valid geodesics and geodesic distances on manifolds. This work opens opportunities for latent-space interpolation, optimal control, and distance computation on differentiable manifolds.
When pretrained language models (LMs) are applied to discriminative tasks such as multiple-choice questions, they place probability mass on vocabulary tokens that aren't among the given answer choices. Spreading probability mass across multiple surface forms with identical meaning (such as "bath" and "bathtub") is thought to cause an underestimation of a model's true performance, referred to as the "surface form competition" (SFC) hypothesis. This has motivated the introduction of various probability normalization methods. However, many core questions remain unanswered. How do we measure SFC? Are there direct ways of reducing it, and does doing so improve task performance? We propose a mathematical formalism for SFC which allows us to quantify and bound its impact for the first time. We identify a simple method for reducing it -- namely, increasing probability mass on the given answer choices by a) including them in the prompt and b) using in-context learning with even just one example. We show this method eliminates the impact of SFC in the majority of instances. Our experiments on three diverse datasets and six LMs reveal several additional surprising findings. For example, both normalization and prompting methods for reducing SFC can be ineffective or even detrimental to task performance for some LMs. We conclude with practical insights for effectively prompting LMs for multiple-choice tasks.
Large language models (LLMs) have achieved widespread success on a variety of in-context few-shot tasks, but this success is typically evaluated via correctness rather than consistency. We argue that self-consistency is an important criteria for valid multi-step reasoning in tasks where the solution is composed of the answers to multiple sub-steps. We propose two types of self-consistency that are particularly important for multi-step reasoning -- hypothetical consistency (a model's ability to predict what its output would be in a hypothetical other context) and compositional consistency (consistency of a model's final outputs when intermediate sub-steps are replaced with the model's outputs for those steps). We demonstrate that multiple variants of the GPT-3/-4 models exhibit poor consistency rates across both types of consistency on a variety of tasks.
The emergence of text-driven motion synthesis technique provides animators with great potential to create efficiently. However, in most cases, textual expressions only contain general and qualitative motion descriptions, while lack fine depiction and sufficient intensity, leading to the synthesized motions that either (a) semantically compliant but uncontrollable over specific pose details, or (b) even deviates from the provided descriptions, bringing animators with undesired cases. In this paper, we propose DiffKFC, a conditional diffusion model for text-driven motion synthesis with KeyFrames Collaborated, enabling realistic generation with collaborative and efficient dual-level control: coarse guidance at semantic level, with only few keyframes for direct and fine-grained depiction down to body posture level. Unlike existing inference-editing diffusion models that incorporate conditions without training, our conditional diffusion model is explicitly trained and can fully exploit correlations among texts, keyframes and the diffused target frames. To preserve the control capability of discrete and sparse keyframes, we customize dilated mask attention modules where only partial valid tokens participate in local-to-global attention, indicated by the dilated keyframe mask. Additionally, we develop a simple yet effective smoothness prior, which steers the generated frames towards seamless keyframe transitions at inference. Extensive experiments show that our model not only achieves state-of-the-art performance in terms of semantic fidelity, but more importantly, is able to satisfy animator requirements through fine-grained guidance without tedious labor.
Despite their impressive performance, large language models (LMs) still struggle with reliably generating complex output structures when not finetuned to follow the required output format exactly. To address this issue, grammar-constrained decoding (GCD) can be used to control the generation of LMs, guaranteeing that the output follows a given structure. Most existing GCD methods are, however, limited to specific tasks, such as parsing or code generation. In this work, we demonstrate that formal grammars can describe the output space for a much wider range of tasks and argue that GCD can serve as a unified framework for structured NLP tasks in general. For increased flexibility, we introduce input-dependent grammars, which allow the grammar to depend on the input and thus enable the generation of different output structures for different inputs. We then empirically demonstrate the power and flexibility of GCD-enhanced LMs on (1) information extraction, (2) entity disambiguation, and (3) constituency parsing. Our results indicate that grammar-constrained LMs substantially outperform unconstrained LMs or even beat task-specific finetuned models. Grammar constraints thus hold great promise for harnessing off-the-shelf LMs for a wide range of structured NLP tasks, especially where training data is scarce or finetuning is expensive. Code and data: https://github.com/epfl-dlab/GCD.
The popularity of Deep Learning (DL), coupled with network traffic visibility reduction due to the increased adoption of HTTPS, QUIC and DNS-SEC, re-ignited interest towards Traffic Classification (TC). However, to tame the dependency from task-specific large labeled datasets we need to find better ways to learn representations that are valid across tasks. In this work we investigate this problem comparing transfer learning, meta-learning and contrastive learning against reference Machine Learning (ML) tree-based and monolithic DL models (16 methods total). Using two publicly available datasets, namely MIRAGE19 (40 classes) and AppClassNet (500 classes), we show that (i) using large datasets we can obtain more general representations, (ii) contrastive learning is the best methodology and (iii) meta-learning the worst one, and (iv) while ML tree-based cannot handle large tasks but fits well small tasks, by means of reusing learned representations, DL methods are reaching tree-based models performance also for small tasks.
We study when the neural tangent kernel (NTK) approximation is valid for training a model with the square loss. In the lazy training setting of Chizat et al. 2019, we show that rescaling the model by a factor of $\alpha = O(T)$ suffices for the NTK approximation to be valid until training time $T$. Our bound is tight and improves on the previous bound of Chizat et al. 2019, which required a larger rescaling factor of $\alpha = O(T^2)$.
Post-selection inference (PoSI) is a statistical technique for obtaining valid confidence intervals and p-values when hypothesis generation and testing use the same source of data. PoSI can be used on a range of popular algorithms including the Lasso. Data carving is a variant of PoSI in which a portion of held out data is combined with the hypothesis generating data at inference time. While data carving has attractive theoretical and empirical properties, existing approaches rely on computationally expensive MCMC methods to carry out inference. This paper's key contribution is to show that pivotal quantities can be constructed for the data carving procedure based on a known parametric distribution. Specifically, when the selection event is characterized by a set of polyhedral constraints on a Gaussian response, data carving will follow the sum of a normal and a truncated normal (SNTN), which is a variant of the truncated bivariate normal distribution. The main impact of this insight is that obtaining exact inference for data carving can be made computationally trivial, since the CDF of the SNTN distribution can be found using the CDF of a standard bivariate normal. A python package sntn has been released to further facilitate the adoption of data carving with PoSI.
In Answer Set Programming (ASP), the user can define declaratively a problem and solve it with efficient solvers; practical applications of ASP are countless and several constraint problems have been successfully solved with ASP. On the other hand, solution time usually grows in a superlinear way (often, exponential) with respect to the size of the instance, which is impractical for large instances. A widely used approach is to split the optimization problem into sub-problems that are solved in sequence, some committing to the values assigned by others, and reconstructing a valid assignment for the whole problem by juxtaposing the solutions of the single sub-problems. On the one hand this approach is much faster, due to the superlinear behavior; on the other hand, it does not provide any guarantee of optimality: committing to the assignment of one sub-problem can rule out the optimal solution from the search space. In other research areas, Logic-Based Benders Decomposition (LBBD) proved effective; in LBBD, the problem is decomposed into a Master Problem (MP) and one or several Sub-Problems (SP). The solution of the MP is passed to the SPs, that can possibly fail. In case of failure, a no-good is returned to the MP, that is solved again with the addition of the new constraint. The solution process is iterated until a valid solution is obtained for all the sub-problems or the MP is proven infeasible. The obtained solution is provably optimal under very mild conditions. In this paper, we apply for the first time LBBD to ASP, exploiting an application in health care as case study. Experimental results show the effectiveness of the approach. We believe that the availability of LBBD can further increase the practical applicability of ASP technologies.