LLMs are frequently fine-tuned or unlearned to adapt to new tasks or eliminate undesirable behaviors.While existing evaluation methods assess performance after such interventions, there remains no general approach for detecting unintended side effects-such as unlearning biology content degrading performance on chemistry tasks, particularly when these effects are unpredictable or emergent.To address this issue, we introduce MNEME, Model diffiNg for Evaluating Mechanistic Effects, a framework for identifying these side effects using sparse model diffing.MNEME compares base and fine-tuned models on out-of-distribution (OOD) data (e.g., The Pile, LMSYS-Chat-1M), without access to fine-tuning data, to isolate behavioral shifts.Applied to five LLMs across three scenarios, WMDP knowledge unlearning, emergent misalignment, and benign finetuning, MNEME achieves up to 95% accuracy in predicting side effects, aligning with known benchmarks and requiring no custom heuristics.Our results demonstrate that sparse probing and diffing offer a scalable and automated lens into fine-tuning-induced model changes, providing practical tools for understanding and managing LLM behavior.1 * MNEME refers to Mnēmosynē, the Greek Titan goddess of memory, whose name derives from the Greek word mnēmē ("memory").1

Deep Generative Models are frequently used to learn continuous representations of complex data distributions by training on a finite number of samples. For any generative model, including pre-trained foundation models with Diffusion or Transformer architectures, generation performance can significantly vary across the learned data manifold. In this paper, we study the local geometry of the learned manifold and its relationship to generation outcomes for a wide range of generative models, including DDPM, Diffusion Transformer (DiT), and Stable Diffusion 1.4. Building on the theory of continuous piecewise-linear (CPWL) generators, we characterize the local geometry in terms of three geometric descriptors - scaling ($\psi$), rank ($\nu$), and complexity/un-smoothness ($\delta$). We provide quantitative and qualitative evidence showing that for a given latent vector, the local descriptors are indicative of post-generation aesthetics, generation diversity, and memorization by the generative model. Finally, we demonstrate that by training a reward model on the 'local scaling' for Stable Diffusion, we can self-improve both generation aesthetics and diversity using geometry sensitive guidance during denoising. Website: https://imtiazhumayun.github.io/generative_geometry.

Generative AI (GAI) is proliferating, and among its many applications are to support creative work (e.g., generating text, images, music) and to enhance accessibility (e.g., captions of images and audio). As GAI evolves, creatives must consider how (or how not) to incorporate these tools into their practices. In this paper, we present interviews at the intersection of these applications. We learned from 10 creatives with disabilities who intentionally use and do not use GAI in and around their creative work. Their mediums ranged from audio engineering to leatherwork, and they collectively experienced a variety of disabilities, from sensory to motor to invisible disabilities. We share cross-cutting themes of their access hacks, how creative practice and access work become entangled, and their perspectives on how GAI should and should not fit into their workflows. In turn, we offer qualities of accessible creativity with responsible AI that can inform future research.

Understanding how communities experience algorithms is necessary to mitigate potential harmful impacts. This paper presents folk theories of text-to-image (T2I) models to enrich understanding of how artist communities experience creative machine learning systems. This research draws on data collected from a workshop with 15 artists from 10 countries who incorporate T2I models in their creative practice. Through reflexive thematic analysis of workshop data, we highlight artist folk theories of T2I use, harm, and harm reduction. Folk theories of use envision T2I models as an artistic medium, a mundane tool, and locate true creativity as rising above model affordances. Theories of harm articulate T2I models as harmed by engineering efforts to eliminate glitches and product policy efforts to limit functionality. Theories of harm-reduction orient towards protecting T2I models for creative practice through transparency and distributed governance. We examine how these theories relate, and conclude by discussing how folk theorization informs responsible AI efforts.

The rise of foundation models holds immense promise for advancing AI, but this progress may amplify existing risks and inequalities, leaving marginalized communities behind. In this position paper, we discuss that disparities towards marginalized communities – performance, representation, privacy, robustness, interpretability and safety – are not isolated concerns but rather interconnected elements of a cascading disparity phenomenon. We contrast foundation models with traditional models and highlight the potential for exacerbated disparity against marginalized communities. Moreover, we emphasize the unique threat of cascading impacts in foundation models, where interconnected disparities can trigger long-lasting negative consequences, specifically to the people on the margin. We define marginalized communities within the machine learning context and explore the multifaceted nature of disparities. We analyze the sources of these disparities, tracing them from data creation, training and deployment procedures to highlight the complex technical and socio-technical landscape. To mitigate the pressing crisis, we conclude with a set of calls to action to mitigate disparity at its source.

Creative ML tools are collaborative systems that afford artistic creativity through their myriad interactive relationships. We propose using "assemblage thinking" to support analyses of creative ML by approaching it as a system in which the elements of people, organizations, culture, practices, and technology constantly influence each other. We model these interactions as "coordinating elements" that give rise to the social and political characteristics of a particular creative ML context, and call attention to three dynamic elements of creative ML whose interactions provide unique context for the social impact a particular system has: people, creative processes, and products. As creative assemblages are highly contextual, we present these as analytical concepts that computing researchers can adapt to better understand the functioning of a particular system or phenomena and identify intervention points to foster desired change. This paper contributes to theorizing interactions with AI in the context of art, and how these interactions shape the production of algorithmic art.

Identifying potential social and ethical risks in emerging machine learning (ML) models and their applications remains challenging. In this work, we applied two well-established safety engineering frameworks (FMEA, STPA) to a case study involving text-to-image models at three stages of the ML product development pipeline: data processing, integration of a T2I model with other models, and use. Results of our analysis demonstrate the safety frameworks – both of which are not designed explicitly examine social and ethical risks – can uncover failure and hazards that pose social and ethical risks. We discovered a broad range of failures and hazards (i.e., functional, social, and ethical) by analyzing interactions (i.e., between different ML models in the product, between the ML product and user, and between development teams) and processes (i.e., preparation of training data or workflows for using an ML service/product). Our findings underscore the value and importance of examining beyond an ML model in examining social and ethical risks, especially when we have minimal information about an ML model.

Understanding the landscape of potential harms from algorithmic systems enables practitioners to better anticipate consequences of the systems they build. It also supports the prospect of incorporating controls to help minimize harms that emerge from the interplay of technologies and social and cultural dynamics. A growing body of scholarship has identified a wide range of harms across different algorithmic technologies. However, computing research and practitioners lack a high level and synthesized overview of harms from algorithmic systems. Based on a scoping review of computing research (n=172), we present an applied taxonomy of sociotechnical harms to support a more systematic surfacing of potential harms in algorithmic systems. The final taxonomy builds on and refers to existing taxonomies, classifications, and terminologies. Five major themes related to sociotechnical harms — representational, allocative, quality-of-service, interpersonal harms, and social system/societal harms — and sub-themes are presented along with a description of these categories. We conclude with a discussion of challenges and opportunities for future research.

Inappropriate design and deployment of machine learning (ML) systems lead to negative downstream social and ethical impacts – described here as social and ethical risks – for users, society, and the environment. Despite the growing need to regulate ML systems, current processes for assessing and mitigating risks are disjointed and inconsistent. We interviewed 30 industry practitioners on their current social and ethical risk management practices and collected their first reactions on adapting safety engineering frameworks into their practice – namely, System Theoretic Process Analysis (STPA) and Failure Mode and Effects Analysis (FMEA). Our findings suggest STPA/FMEA can provide an appropriate structure for social and ethical risk assessment and mitigation processes. However, we also find nontrivial challenges in integrating such frameworks in the fast-paced culture of the ML industry. We call on the CHI community to strengthen existing frameworks and assess their efficacy, ensuring that ML systems are safer for all people.

Instance segmentation methods often require costly per-pixel labels. We propose a method called WISE-Net that only requires point-level annotations. During training, the model only has access to a single pixel label per object, yet the task is to output full segmentation masks. To address this challenge, we construct a network with two branches: (1) a 10-calization network (L-Net) that predicts the location of each object; and (2) an embedding network (E-Net) that learns an embedding space where pixels of the same object are close. The segmentation masks for the located objects are obtained by grouping pixels with similar embeddings. We evaluate our approach on PASCAL VOC, COCO, KITTI and CityScapes datasets. The experiments show that our method (1) obtains competitive results compared to fully-supervised methods in certain scenarios; (2) outperforms fully-and weakly-supervised methods with a fixed annotation budget; and (3) establishes a first strong baseline for instance segmentation with point-level supervision.

Neural networks are prone to adversarial attacks. In general, such attacks deteriorate the quality of the input by either slightly modifying most of its pixels, or by occluding it with a patch. In this paper, we propose a method that keeps the image unchanged and only adds an adversarial framing on the border of the image. We show empirically that our method is able to successfully attack state-of-theart methods on both image and video classification problems. Notably, the proposed method results in a universal attack which is very fast at test time. Source code can be found at github.com/zajaczajac/adv_framing.

Single-view 3D shape reconstruction is an important but challenging problem, mainly for two reasons. First, as shape annotation is very expensive to acquire, current methods rely on synthetic data, in which ground-truth 3D annotation is easy to obtain. However, this results in domain adaptation problem when applied to natural images. The second challenge is that there are multiple shapes that can explain a given 2D image. In this paper, we propose a framework to improve over these challenges using adversarial training. On one hand, we impose domain confusion between natural and synthetic image representations to reduce the distribution gap. On the other hand, we impose the reconstruction to be `realistic' by forcing it to lie on a (learned) manifold of realistic object shapes. Our experiments show that these constraints improve performance by a large margin over baseline reconstruction models. We achieve results competitive with the state of the art with a much simpler architecture.

For embodied agents to infer representations of the underlying 3D physical world they inhabit, they should efficiently combine multisensory cues from numerous trials, e.g., by looking at and touching objects. Despite its importance, multisensory 3D scene representation learning has received less attention compared to the unimodal setting. In this paper, we propose the Generative Multisensory Network (GMN) for learning latent representations of 3D scenes which are partially observable through multiple sensory modalities. We also introduce a novel method, called the Amortized Product-of-Experts, to improve the computational efficiency and the robustness to unseen combinations of modalities at test time. Experimental results demonstrate that the proposed model can efficiently infer robust modality-invariant 3D-scene representations from arbitrary combinations of modalities and perform accurate cross-modal generation. To perform this exploration we have also developed a novel multi-sensory simulation environment for embodied agents.

Metric-based meta-learning techniques have successfully been applied to few-shot classification problems. In this paper, we propose to leverage cross-modal information to enhance metric-based few-shot learning methods. Visual and semantic feature spaces have different structures by definition. For certain concepts, visual features might be richer and more discriminative than text ones. While for others, the inverse might be true. Moreover, when the support from visual information is limited in image classification, semantic representations (learned from unsupervised text corpora) can provide strong prior knowledge and context to help learning. Based on these two intuitions, we propose a mechanism that can adaptively combine information from both modalities according to new image categories to be learned. Through a series of experiments, we show that by this adaptive combination of the two modalities, our model outperforms current uni-modality few-shot learning methods and modality-alignment methods by a large margin on all benchmarks and few-shot scenarios tested. Experiments also show that our model can effectively adjust its focus on the two modalities. The improvement in performance is particularly large when the number of shots is very small.

Classical Bag-of-Words methods represent videos by modeling the variation of local visual descriptors throughout the video. In this approach they mix variation in time and space indiscriminately while these dimensions are fundamentally different. Therefore, in this paper we present a novel method for video representation which explicitly captures temporal variation over time. We do this by first creating frame-based features using standard Bag-of-Words techniques. To model the variation in time over these frame-based features, we introduce Hard and Soft Cluster Encoding, novel techniques to model variation inspired by the Fisher Kernel [1] and VLAD [2]. Results on the Rochester ADL [3] and Blip10k [4] datasets show that our method yields improvements of respectively 6.6% and 7.4% over our baselines. On Blip10k we outperform the state-of-the-art by 3.6% when using only visual features.

Motivated by applications in areas such as patient monitoring, tele-rehabilitation and ambient assisted living, analyzing activities of daily living is an active research topic in computer vision and image processing. In this paper we address the problem of everyday activity recognition from unlabeled data proposing a novel multi-task clustering (MTC) approach. Our intuition is that, when analyzing activities of daily living, we can take advantage of the fact that people tend to perform the same actions in the same environment (e.g. people working in an office environment use to read and write documents). Thus, even if labels are not available, information about typical activities can be exploited in the learning process. Arguing that the tasks of recognizing activities of specific individuals are related, we resort on multi-task learning and rather than clustering the data of each individual separately, we also look for clustering results which are coherent among related tasks. Extensive experimental results show that our method outperforms several state-of-the-art approaches by up to 11% on the Rochester activities of daily living dataset.

The current state-of-the-art in Video Classification is based on Bag-of-Words using local visual descriptors. Most commonly these are Histogram of Oriented Gradient (HOG) and Histogram of Optical Flow (HOF) descriptors. While such system is very powerful for classification, it is also computationally expensive. This paper addresses the problem of computational efficiency. Specifically: (1) We propose several speed-ups for densely sampled HOG and HOF descriptors and release Matlab code. (2) We investigate the trade-off between accuracy and computational efficiency of descriptors in terms of frame sampling rate and type of Optical Flow method. (3) We investigate the trade-off between accuracy and computational efficiency for the video representation, using either a k-means or hierarchical k-means based visual vocabulary, a Random Forest based vocabulary or the Fisher kernel.

In video global features are often used for reasons of computational efficiency, where each global feature captures information of a single video frame. But frames in video change over time, so an important question is: how can we meaningfully aggregate frame-based features in order to preserve the variation in time? In this paper we propose to use the Fisher Kernel to capture variation in time in video. While in this approach the temporal order is lost, it captures both subtle variation in time such as the ones caused by a moving bicycle and drastic variations in time such as the changing of shots in a documentary. Our work should not be confused with a Bag of Local Visual Features approach, where one captures the visual variation of local features in both time and space indiscriminately. Instead, each feature measures a complete frame hence we capture variation in time only.