Eric Yeats

Eric Yeats, Frank Liu, David Womble et al.

We present a self-supervised method to disentangle factors of variation in high-dimensional data that does not rely on prior knowledge of the underlying variation profile (e.g., no assumptions on the number or distribution of the individual latent variables to be extracted). In this method which we call NashAE, high-dimensional feature disentanglement is accomplished in the low-dimensional latent space of a standard autoencoder (AE) by promoting the discrepancy between each encoding element and information of the element recovered from all other encoding elements. Disentanglement is promoted efficiently by framing this as a minmax game between the AE and an ensemble of regression networks which each provide an estimate of an element conditioned on an observation of all other elements. We quantitatively compare our approach with leading disentanglement methods using existing disentanglement metrics. Furthermore, we show that NashAE has increased reliability and increased capacity to capture salient data characteristics in the learned latent representation.

Eric Yeats, Frank Liu, Hai Li

Disentangled learning representations have promising utility in many applications, but they currently suffer from serious reliability issues. We present Gaussian Channel Autoencoder (GCAE), a method which achieves reliable disentanglement via flexible density estimation of the latent space. GCAE avoids the curse of dimensionality of density estimation by disentangling subsets of its latent space with the Dual Total Correlation (DTC) metric, thereby representing its high-dimensional latent joint distribution as a collection of many low-dimensional conditional distributions. In our experiments, GCAE achieves highly competitive and reliable disentanglement scores compared with state-of-the-art baselines.

Jingyang Zhang, Jingwei Sun, Eric Yeats et al.

The problem of pre-training data detection for large language models (LLMs) has received growing attention due to its implications in critical issues like copyright violation and test data contamination. Despite improved performance, existing methods (including the state-of-the-art, Min-K%) are mostly developed upon simple heuristics and lack solid, reasonable foundations. In this work, we propose a novel and theoretically motivated methodology for pre-training data detection, named Min-K%++. Specifically, we present a key insight that training samples tend to be local maxima of the modeled distribution along each input dimension through maximum likelihood training, which in turn allow us to insightfully translate the problem into identification of local maxima. Then, we design our method accordingly that works under the discrete distribution modeled by LLMs, whose core idea is to determine whether the input forms a mode or has relatively high probability under the conditional categorical distribution. Empirically, the proposed method achieves new SOTA performance across multiple settings. On the WikiMIA benchmark, Min-K%++ outperforms the runner-up by 6.2% to 10.5% in detection AUROC averaged over five models. On the more challenging MIMIR benchmark, it consistently improves upon reference-free methods while performing on par with reference-based method that requires an extra reference model.

Eric Yeats, Cameron Darwin, Frank Liu et al.

Quantification of the number of variables needed to locally explain complex data is often the first step to better understanding it. Existing techniques from intrinsic dimension estimation leverage statistical models to glean this information from samples within a neighborhood. However, existing methods often rely on well-picked hyperparameters and ample data as manifold dimension and curvature increases. Leveraging insight into the fixed point of the score matching objective as the score map is regularized by its Dirichlet energy, we show that it is possible to retrieve the topological dimension of the manifold learned by the score map. We then introduce a novel method to measure the learned manifold's topological dimension (i.e., local intrinsic dimension) using adversarial attacks, thereby generating useful interpretations of the learned manifold.

Eric Yeats, Darryl Hannan, Henry Kvinge et al.

Machine unlearning (MU) is a promising cost-effective method to cleanse undesired information (generated concepts, biases, or patterns) from foundational diffusion models. While MU is orders of magnitude less costly than retraining a diffusion model without the undesired information, it can be challenging and labor-intensive to prove that the information has been fully removed from the model. Moreover, MU can damage diffusion model performance on surrounding concepts that one would like to retain, making it unclear if the diffusion model is still fit for deployment. We introduce autoeval-dmun, an automated tool which leverages (vision-) language models to thoroughly assess unlearning in diffusion models. Given a target concept, autoeval-dmun extracts structured, relevant world knowledge from the language model to identify nearby concepts which are likely damaged by unlearning and to circumvent unlearning with adversarial prompts. We use our automated tool to evaluate popular diffusion model unlearning methods, revealing that language models (1) impose semantic orderings of nearby concepts which correlate well with unlearning damage and (2) effectively circumvent unlearning with synthetic adversarial prompts.

Eric Yeats, John Buckheit, Sarah Scullen et al.

Hallucination remains a barrier to deploying generative models in high-consequence applications. This is especially true in cases where external ground truth is not readily available to validate model outputs. This situation has motivated the study of geometric signals in the internal state of an LLM that are predictive of hallucination and require limited external knowledge. Given that there are a range of factors that can lead model output to be called a hallucination (e.g., irrelevance vs incoherence), in this paper we ask what specific properties of a hallucination these geometric statistics actually capture. To assess this, we generate a synthetic dataset which varies distinct properties of output associated with hallucination. This includes output correctness, confidence, relevance, coherence, and completeness. We find that different geometric statistics capture different types of hallucinations. Along the way we show that many existing geometric detection methods have substantial sensitivity to shifts in task domain (e.g., math questions vs. history questions). Motivated by this, we introduce a simple normalization method to mitigate the effect of domain shift on geometric statistics, leading to AUROC gains of +34 points in multi-domain settings.

Eric Yeats, Darryl Hannan, Wilson Fearn et al.

Score-based generative models require guidance in order to generate plausible, on-manifold samples. The most popular guidance method, Classifier-Free Guidance (CFG), is only applicable in settings with labeled data and requires training an additional unconditional score-based model. More recently, Auto-Guidance adopts a smaller, less capable version of the original model to guide generation. While each method effectively promotes the fidelity of generated data, each requires labeled data or the training of additional models, making it challenging to guide score-based models when (labeled) training data are not available or training new models is not feasible. We make the surprising discovery that the positive curvature of log density estimates in saddle regions provides strong guidance for score-based models. Motivated by this, we develop saddle-free guidance (SFG) which maintains estimates of maximal positive curvature of the log density to guide individual score-based models. SFG has the same computational cost of classifier-free guidance, does not require additional training, and works with off-the-shelf diffusion and flow matching models. Our experiments indicate that SFG achieves state-of-the-art FID and FD-DINOv2 metrics in single-model unconditional ImageNet-512 generation. When SFG is combined with Auto-Guidance, its unconditional samples achieve general state-of-the-art in FD-DINOv2 score. Our experiments with FLUX.1-dev and Stable Diffusion v3.5 indicate that SFG boosts the diversity of output images compared to CFG while maintaining excellent prompt adherence and image fidelity.

Eric Yeats, Aaron Jacobson, Darryl Hannan et al.

The local intrinsic dimension (LID) of data is a fundamental quantity in signal processing and learning theory, but quantifying the LID of high-dimensional, complex data has been a historically challenging task. Recent works have discovered that diffusion models capture the LID of data through the spectra of their score estimates and through the rate of change of their density estimates under various noise perturbations. While these methods can accurately quantify LID, they require either many forward passes of the diffusion model or use of gradient computation, limiting their applicability in compute- and memory-constrained scenarios. We show that the LID is a lower bound on the denoising score matching loss, motivating use of the denoising score matching loss as a LID estimator. Moreover, we show that the equivalent implicit score matching loss also approximates LID via the normal dimension and is closely related to a recent LID estimator, FLIPD. Our experiments on a manifold benchmark and with Stable Diffusion 3.5 indicate that the denoising score matching loss is a highly competitive and scalable LID estimator, achieving superior accuracy and memory footprint under increasing problem size and quantization level.

Eric Yeats, Cameron Darwin, Eduardo Ortega et al.

We leverage diffusion models to study the robustness-performance tradeoff of robust classifiers. Our approach introduces a simple, pretrained diffusion method to generate low-norm counterfactual examples (CEs): semantically altered data which results in different true class membership. We report that the confidence and accuracy of robust models on their clean training data are associated with the proximity of the data to their CEs. Moreover, robust models perform very poorly when evaluated on the CEs directly, as they become increasingly invariant to the low-norm, semantic changes brought by CEs. The results indicate a significant overlap between non-robust and semantic features, countering the common assumption that non-robust features are not interpretable.