Luca Zappella

Borja Rodríguez-Gálvez, Arno Blaas, Pau Rodríguez et al. · apple-ml

The mechanisms behind the success of multi-view self-supervised learning (MVSSL) are not yet fully understood. Contrastive MVSSL methods have been studied through the lens of InfoNCE, a lower bound of the Mutual Information (MI). However, the relation between other MVSSL methods and MI remains unclear. We consider a different lower bound on the MI consisting of an entropy and a reconstruction term (ER), and analyze the main MVSSL families through its lens. Through this ER bound, we show that clustering-based methods such as DeepCluster and SwAV maximize the MI. We also re-interpret the mechanisms of distillation-based approaches such as BYOL and DINO, showing that they explicitly maximize the reconstruction term and implicitly encourage a stable entropy, and we confirm this empirically. We show that replacing the objectives of common MVSSL methods with this ER bound achieves competitive performance, while making them stable when training with smaller batch sizes or smaller exponential moving average (EMA) coefficients. Github repo: https://github.com/apple/ml-entropy-reconstruction.

Aspen Hopkins, Fred Hohman, Luca Zappella et al. · apple-ml, cmu

Lack of diversity in data collection has caused significant failures in machine learning (ML) applications. While ML developers perform post-collection interventions, these are time intensive and rarely comprehensive. Thus, new methods to track & manage data collection, iteration, and model training are necessary for evaluating whether datasets reflect real world variability. We present designing data, an iterative approach to data collection connecting HCI concepts with ML techniques. Our process includes (1) Pre-Collection Planning, to reflexively prompt and document expected data distributions; (2) Collection Monitoring, to systematically encourage sampling diversity; and (3) Data Familiarity, to identify samples that are unfamiliar to a model using density estimation. We apply designing data to a data collection and modeling task. We find models trained on ''designed'' datasets generalize better across intersectional groups than those trained on similarly sized but less targeted datasets, and that data familiarity is effective for debugging datasets.

Xavier Suau, Pieter Delobelle, Katherine Metcalf et al.

An important issue with Large Language Models (LLMs) is their undesired ability to generate toxic language. In this work, we show that the neurons responsible for toxicity can be determined by their power to discriminate toxic sentences, and that toxic language can be mitigated by reducing their activation levels proportionally to this power. We propose AUROC adaptation (AurA), an intervention that can be applied to any pre-trained LLM to mitigate toxicity. As the intervention is proportional to the ability of each neuron to discriminate toxic content, it is free of any model-dependent hyperparameters. We show that AurA can achieve up to $2.2 \times$ reduction in toxicity with only a $0.72$ perplexity increase. We also show that AurA is effective with models of different scale (from 1.5B to 40B parameters), and its effectiveness in mitigating toxic language, while preserving common-sense zero-shot abilities, holds across all scales. AurA can be combined with pre-prompting strategies, boosting its average mitigation potential from $1.28\times$ to $2.35\times$. Moreover, AurA can counteract adversarial pre-prompts that maliciously elicit toxic content, making it an effective method for deploying safer and less toxic models.

Louis Bethune, Victor Turrisi, Bruno Kacper Mlodozeniec et al. · apple-ml, berkeley

Discrete diffusion models have emerged as strong alternatives to autoregressive language models, with recent work initializing and fine-tuning a base unimodal model for bimodal generation. Diverging from previous approaches, we introduce the first tri-modal masked diffusion model pretrained from scratch on text, image-text, and audio-text data. We systematically analyze multimodal scaling laws, modality mixing ratios, noise schedules, and batch-size effects, and we provide optimized inference sampling defaults. Our batch-size analysis yields a novel stochastic differential equation (SDE)-based reparameterization that eliminates the need for tuning the optimal batch size as reported in recent work. This reparameterization decouples the physical batch size, often chosen based on compute constraints (GPU saturation, FLOP efficiency, wall-clock time), from the logical batch size, chosen to balance gradient variance during stochastic optimization. Finally, we pretrain a preliminary 3B-parameter tri-modal model on 6.4T tokens, demonstrating the capabilities of a unified design and achieving strong results in text generation, text-to-image tasks, and text-to-speech tasks. Our work represents the largest-scale systematic open study of multimodal discrete diffusion models conducted to date, providing insights into scaling behaviors across multiple modalities.

Miguel Sarabia, Elena Menyaylenko, Alessandro Toso et al.

We present Spatial LibriSpeech, a spatial audio dataset with over 650 hours of 19-channel audio, first-order ambisonics, and optional distractor noise. Spatial LibriSpeech is designed for machine learning model training, and it includes labels for source position, speaking direction, room acoustics and geometry. Spatial LibriSpeech is generated by augmenting LibriSpeech samples with 200k+ simulated acoustic conditions across 8k+ synthetic rooms. To demonstrate the utility of our dataset, we train models on four spatial audio tasks, resulting in a median absolute error of 6.60° on 3D source localization, 0.43m on distance, 90.66ms on T30, and 2.74dB on DRR estimation. We show that the same models generalize well to widely-used evaluation datasets, e.g., obtaining a median absolute error of 12.43° on 3D source localization on TUT Sound Events 2018, and 157.32ms on T30 estimation on ACE Challenge.

Jonathan Crabbé, Pau Rodríguez, Vaishaal Shankar et al.

What distinguishes robust models from non-robust ones? While for ImageNet distribution shifts it has been shown that such differences in robustness can be traced back predominantly to differences in training data, so far it is not known what that translates to in terms of what the model has learned. In this work, we bridge this gap by probing the representation spaces of 16 robust zero-shot CLIP vision encoders with various backbones (ResNets and ViTs) and pretraining sets (OpenAI, LAION-400M, LAION-2B, YFCC15M, CC12M and {DataComp}), and comparing them to the representation spaces of less robust models with identical backbones, but different (pre)training sets or objectives (CLIP pretraining on ImageNet-Captions, and supervised training or finetuning on ImageNet).Through this analysis, we generate three novel insights. Firstly, we detect the presence of outlier features in robust zero-shot CLIP vision encoders, which to the best of our knowledge is the first time these are observed in non-language and non-transformer models. Secondly, we find the existence of outlier features to be an indication of ImageNet shift robustness in models, since we only find them in robust models in our analysis. Lastly, we also investigate the number of unique encoded concepts in the representation space and find zero-shot CLIP models to encode a higher number of unique concepts in their representation space. However, we do not find this to be an indicator of ImageNet shift robustness and hypothesize that it is rather related to the language supervision. Since the presence of outlier features can be detected without access to any data from shifted datasets, we believe that they could be a useful tool for practitioners to get a feeling for the distribution shift robustness of a pretrained model during deployment.

Abhinav Moudgil, Ningyuan Huang, Eeshan Gunesh Dhekane et al. · apple-ml, mila

State Space Models (SSMs) such as Mamba have become a popular alternative to Transformer models, due to their reduced memory consumption and higher throughput at generation compared to their Attention-based counterparts. On the other hand, the community has built up a considerable body of knowledge on how to train Transformers, and many pretrained Transformer models are readily available. To facilitate the adoption of SSMs while leveraging existing pretrained Transformers, we aim to identify an effective recipe to distill an Attention-based model into a Mamba-like architecture. In prior work on cross-architecture distillation, however, it has been shown that a naïve distillation procedure from Transformers to Mamba fails to preserve the original teacher performance, a limitation often overcome with hybrid solutions combining Attention and SSM blocks. The key argument from our work is that, by equipping Mamba with a principled initialization, we can recover an overall better recipe for cross-architectural distillation. To this end, we propose a principled two-stage approach: first, we distill knowledge from a traditional Transformer into a linearized version of Attention, using an adaptation of the kernel trick. Then, we distill the linearized version into an adapted Mamba model that does not use any Attention block. Overall, the distilled Mamba model is able to preserve the original Pythia-1B Transformer performance in downstream tasks, maintaining a perplexity of 14.11 close to the teacher's 13.86. To show the efficacy of our recipe, we conduct thorough ablations at 1B scale with 10B tokens varying sequence mixer architecture, scaling analysis on model sizes and total distillation tokens, and a sensitivity analysis on tokens allocation between stages.

Xavier Suau, Federico Danieli, T. Anderson Keller et al.

Multiview Self-Supervised Learning (MSSL) is based on learning invariances with respect to a set of input transformations. However, invariance partially or totally removes transformation-related information from the representations, which might harm performance for specific downstream tasks that require such information. We propose 2D strUctured and EquivarianT representations (coined DUET), which are 2d representations organized in a matrix structure, and equivariant with respect to transformations acting on the input data. DUET representations maintain information about an input transformation, while remaining semantically expressive. Compared to SimCLR (Chen et al., 2020) (unstructured and invariant) and ESSL (Dangovski et al., 2022) (unstructured and equivariant), the structured and equivariant nature of DUET representations enables controlled generation with lower reconstruction error, while controllability is not possible with SimCLR or ESSL. DUET also achieves higher accuracy for several discriminative tasks, and improves transfer learning.

Federico Danieli, Miguel Sarabia, Xavier Suau et al.

Parallelization techniques have become ubiquitous for accelerating inference and training of deep neural networks. Despite this, several operations are still performed in a sequential manner. For instance, the forward and backward passes are executed layer-by-layer, and the output of diffusion models is produced by applying a sequence of denoising steps. This sequential approach results in a computational cost proportional to the number of steps involved, presenting a potential bottleneck as the number of steps increases. In this work, we introduce DeepPCR, a novel algorithm which parallelizes typically sequential operations in order to speed up inference and training of neural networks. DeepPCR is based on interpreting a sequence of $L$ steps as the solution of a specific system of equations, which we recover using the Parallel Cyclic Reduction algorithm. This reduces the complexity of computing the sequential operations from $\mathcal{O}(L)$ to $\mathcal{O}(\log_2L)$, thus yielding a speedup for large $L$. To verify the theoretical lower complexity of the algorithm, and to identify regimes for speedup, we test the effectiveness of DeepPCR in parallelizing the forward and backward pass in multi-layer perceptrons, and reach speedups of up to $30\times$ for the forward and $200\times$ for the backward pass. We additionally showcase the flexibility of DeepPCR by parallelizing training of ResNets with as many as 1024 layers, and generation in diffusion models, enabling up to $7\times$ faster training and $11\times$ faster generation, respectively, when compared to the sequential approach.

T. Anderson Keller, Xavier Suau, Luca Zappella

In this work, we observe that many existing self-supervised learning algorithms can be both unified and generalized when seen through the lens of equivariant representations. Specifically, we introduce a general framework we call Homomorphic Self-Supervised Learning, and theoretically show how it may subsume the use of input-augmentations provided an augmentation-homomorphic feature extractor. We validate this theory experimentally for simple augmentations, demonstrate how the framework fails when representational structure is removed, and further empirically explore how the parameters of this framework relate to those of traditional augmentation-based self-supervised learning. We conclude with a discussion of the potential benefits afforded by this new perspective on self-supervised learning.

Nico Lingg, Miguel Sarabia, Luca Zappella et al.

Human skeleton point clouds are commonly used to automatically classify and predict the behaviour of others. In this paper, we use a contrastive self-supervised learning method, SimCLR, to learn representations that capture the semantics of skeleton point clouds. This work focuses on systematically evaluating the effects that different algorithmic decisions (including augmentations, dataset partitioning and backbone architecture) have on the learned skeleton representations. To pre-train the representations, we normalise six existing datasets to obtain more than 40 million skeleton frames. We evaluate the quality of the learned representations with three downstream tasks: skeleton reconstruction, motion prediction, and activity classification. Our results demonstrate the importance of 1) combining spatial and temporal augmentations, 2) including additional datasets for encoder training, and 3) and using a graph neural network as an encoder.

Emily Cheng, Carmen Amo Alonso, Federico Danieli et al.

As generative models become ubiquitous, there is a critical need for fine-grained control over the generation process. Yet, while controlled generation methods from prompting to fine-tuning proliferate, a fundamental question remains unanswered: are these models truly controllable in the first place? In this work, we provide a theoretical framework to formally answer this question. Framing human-model interaction as a control process, we propose a novel algorithm to estimate the controllable sets of models in a dialogue setting. Notably, we provide formal guarantees on the estimation error as a function of sample complexity: we derive probably-approximately correct bounds for controllable set estimates that are distribution-free, employ no assumptions except for output boundedness, and work for any black-box nonlinear control system (i.e., any generative model). We empirically demonstrate the theoretical framework on different tasks in controlling dialogue processes, for both language models and text-to-image generation. Our results show that model controllability is surprisingly fragile and highly dependent on the experimental setting. This highlights the need for rigorous controllability analysis, shifting the focus from simply attempting control to first understanding its fundamental limits.

Federico Danieli, Pau Rodriguez, Miguel Sarabia et al.

Recurrent Neural Networks (RNNs) laid the foundation for sequence modeling, but their intrinsic sequential nature restricts parallel computation, creating a fundamental barrier to scaling. This has led to the dominance of parallelizable architectures like Transformers and, more recently, State Space Models (SSMs). While SSMs achieve efficient parallelization through structured linear recurrences, this linearity constraint limits their expressive power and precludes modeling complex, nonlinear sequence-wise dependencies. To address this, we present ParaRNN, a framework that breaks the sequence-parallelization barrier for nonlinear RNNs. Building on prior work, we cast the sequence of nonlinear recurrence relationships as a single system of equations, which we solve in parallel using Newton's iterations combined with custom parallel reductions. Our implementation achieves speedups of up to 665x over naive sequential application, allowing training nonlinear RNNs at unprecedented scales. To showcase this, we apply ParaRNN to adaptations of LSTM and GRU architectures, successfully training models of 7B parameters that attain perplexity comparable to similarly-sized Transformers and Mamba2 architectures. To accelerate research in efficient sequence modeling, we release the ParaRNN codebase as an open-source framework for automatic training-parallelization of nonlinear RNNs, enabling researchers and practitioners to explore new nonlinear RNN models at scale.

Yinong Oliver Wang, Nivedha Sivakumar, Falaah Arif Khan et al.

The recent rapid adoption of large language models (LLMs) highlights the critical need for benchmarking their fairness. Conventional fairness metrics, which focus on discrete accuracy-based evaluations (i.e., prediction correctness), fail to capture the implicit impact of model uncertainty (e.g., higher model confidence about one group over another despite similar accuracy). To address this limitation, we propose an uncertainty-aware fairness metric, UCerF, to enable a fine-grained evaluation of model fairness that is more reflective of the internal bias in model decisions compared to conventional fairness measures. Furthermore, observing data size, diversity, and clarity issues in current datasets, we introduce a new gender-occupation fairness evaluation dataset with 31,756 samples for co-reference resolution, offering a more diverse and suitable dataset for evaluating modern LLMs. We establish a benchmark, using our metric and dataset, and apply it to evaluate the behavior of ten open-source LLMs. For example, Mistral-7B exhibits suboptimal fairness due to high confidence in incorrect predictions, a detail overlooked by Equalized Odds but captured by UCerF. Overall, our proposed LLM benchmark, which evaluates fairness with uncertainty awareness, paves the way for developing more transparent and accountable AI systems.

Valentino Maiorca, Eleonora Gualdoni, Xavier Suau et al.

As foundation models grow in capability, the ability to efficiently and reliably control their behavior becomes critical. Fine-tuning these models can be costly, and while prompting can be practical for controllability, it remains fragile due to models' high sensitivity to exact prompt wording and structure. This brittleness has driven interest in activation steering techniques that offer more stable and predictable control over model behavior. However, existing activation steering methods require per-concept optimization, which makes them ill-suited to deployment scenarios where the concept set is large, evolving, or only specified at request time: each new concept incurs at least minutes of optimization on the target model. We propose HyperTransport, a hypernetwork framework that amortizes this cost by mapping embeddings from a pretrained encoder (CLIP in our instantiation) directly to intervention parameters, trained end-to-end using an optimal transport loss. Once trained, HyperTransport produces each new intervention in a single hypernetwork forward pass, 3600-7000x faster than per-concept fitting. On concepts unseen during training, it matches the strongest per-concept baselines at inducing the target concept. By decoupling concept representation from intervention prediction, HyperTransport combines three capabilities that no existing approach offers as a set: amortized steering for open-ended concept sets, continuous interpretable strength control, and cross-modal conditioning where reference images can directly steer text-based generation. We validate HyperTransport on DMD2 and Nitro-1-PixArt across 167 held-out test concepts via CLIP-based metrics, a VLM-as-a-judge evaluation, and a user study. In pairwise comparisons, both human and VLM judges prefer HyperTransport over prompting ~2x as often.

Pau Rodriguez, Arno Blaas, Michal Klein et al.

The increasing capabilities of large generative models and their ever more widespread deployment have raised concerns about their reliability, safety, and potential misuse. To address these issues, recent works have proposed to control model generation by steering model activations in order to effectively induce or prevent the emergence of concepts or behaviors in the generated output. In this paper we introduce Activation Transport (AcT), a general framework to steer activations guided by optimal transport theory that generalizes many previous activation-steering works. AcT is modality-agnostic and provides fine-grained control over the model behavior with negligible computational overhead, while minimally impacting model abilities. We experimentally show the effectiveness and versatility of our approach by addressing key challenges in large language models (LLMs) and text-to-image diffusion models (T2Is). For LLMs, we show that AcT can effectively mitigate toxicity, induce arbitrary concepts, and increase their truthfulness. In T2Is, we show how AcT enables fine-grained style control and concept negation.

Zihuiwen Ye, Lukas Aichberger, Michael Kirchhof et al.

Large Language Models (LLMs) are increasingly deployed to autonomously solve real-world tasks. A key ingredient for this is the LLM Function-Calling paradigm, a widely used approach for equipping LLMs with tool-use capabilities. However, an LLM calling functions incorrectly can have severe implications, especially when their effects are irreversible, e.g., transferring money or deleting data. Hence, it is of paramount importance to consider the LLM's confidence that a function call solves the task correctly prior to executing it. Uncertainty Quantification (UQ) methods can be used to quantify this confidence and prevent potentially incorrect function calls. In this work, we present what is, to our knowledge, the first evaluation of UQ methods for LLM Function-Calling (FC). While multi-sample UQ methods, such as Semantic Entropy, show strong performance for natural language Q&A tasks, we find that in the FC setting, it offers no clear advantage over simple single-sample UQ methods. Additionally, we find that the particularities of FC outputs can be leveraged to improve the performance of existing UQ methods in this setting. Specifically, multi-sample UQ methods benefit from clustering FC outputs based on their abstract syntax tree parsing, while single-sample UQ methods can be improved by selecting only semantically meaningful tokens when calculating logit-based uncertainty scores.

Andrea Santilli, Adam Golinski, Michael Kirchhof et al.

Uncertainty Quantification (UQ) in Language Models (LMs) is key to improving their safety and reliability. Evaluations often use metrics like AUROC to assess how well UQ methods (e.g., negative sequence probabilities) correlate with task correctness functions (e.g., ROUGE-L). We show that mutual biases--when both UQ methods and correctness functions are biased by the same factors--systematically distort evaluation. First, we formally prove that any mutual bias non-randomly skews AUROC rankings, compromising benchmark integrity. Second, we confirm this happens empirically by testing 7 widely used correctness functions, from lexical-based and embedding-based metrics to LM-as-a-judge approaches, across 4 datasets x 4 models x 8 UQ methods. Our analysis shows that length biases in correctness functions distort UQ assessments by interacting with length biases in UQ methods. We identify LM-as-a-judge methods as the least length-biased, offering a promising path for a fairer UQ evaluation.

Natalie Mackraz, Nivedha Sivakumar, Samira Khorshidi et al.

Large language models (LLMs) are increasingly being adapted to achieve task-specificity for deployment in real-world decision systems. Several previous works have investigated the bias transfer hypothesis (BTH) by studying the effect of the fine-tuning adaptation strategy on model fairness to find that fairness in pre-trained masked language models have limited effect on the fairness of models when adapted using fine-tuning. In this work, we expand the study of BTH to causal models under prompt adaptations, as prompting is an accessible, and compute-efficient way to deploy models in real-world systems. In contrast to previous works, we establish that intrinsic biases in pre-trained Mistral, Falcon and Llama models are strongly correlated (rho >= 0.94) with biases when the same models are zero- and few-shot prompted, using a pronoun co-reference resolution task. Further, we find that bias transfer remains strongly correlated even when LLMs are specifically prompted to exhibit fair or biased behavior (rho >= 0.92), and few-shot length and stereotypical composition are varied (rho >= 0.97). Our findings highlight the importance of ensuring fairness in pre-trained LLMs, especially when they are later used to perform downstream tasks via prompt adaptation.

Nivedha Sivakumar, Natalie Mackraz, Samira Khorshidi et al.

A dangerous assumption that can be made from prior work on the bias transfer hypothesis (BTH) is that biases do not transfer from pre-trained large language models (LLMs) to adapted models. We invalidate this assumption by studying the BTH in causal models under prompt adaptations, as prompting is an extremely popular and accessible adaptation strategy used in real-world applications. In contrast to prior work, we find that biases can transfer through prompting and that popular prompt-based mitigation methods do not consistently prevent biases from transferring. Specifically, the correlation between intrinsic biases and those after prompt adaptation remain moderate to strong across demographics and tasks -- for example, gender (rho >= 0.94) in co-reference resolution, and age (rho >= 0.98) and religion (rho >= 0.69) in question answering. Further, we find that biases remain strongly correlated when varying few-shot composition parameters, such as sample size, stereotypical content, occupational distribution and representational balance (rho >= 0.90). We evaluate several prompt-based debiasing strategies and find that different approaches have distinct strengths, but none consistently reduce bias transfer across models, tasks or demographics. These results demonstrate that correcting bias, and potentially improving reasoning ability, in intrinsic models may prevent propagation of biases to downstream tasks.

Ningyuan Huang, Miguel Sarabia, Abhinav Moudgil et al.

State-Space Models (SSMs), and particularly Mamba, have recently emerged as a promising alternative to Transformers. Mamba introduces input selectivity to its SSM layer (S6) and incorporates convolution and gating into its block definition. While these modifications do improve Mamba's performance over its SSM predecessors, it remains largely unclear how Mamba leverages the additional functionalities provided by input selectivity, and how these interact with the other operations in the Mamba architecture. In this work, we demystify the role of input selectivity in Mamba, investigating its impact on function approximation power, long-term memorization, and associative recall capabilities. In particular: (i) we prove that the S6 layer of Mamba can represent projections onto Haar wavelets, providing an edge over its Diagonal SSM (S4D) predecessor in approximating discontinuous functions commonly arising in practice; (ii) we show how the S6 layer can dynamically counteract memory decay; (iii) we provide analytical solutions to the MQAR associative recall task using the Mamba architecture with different mixers -- Mamba, Mamba-2, and S4D. We demonstrate the tightness of our theoretical constructions with empirical results on concrete tasks. Our findings offer a mechanistic understanding of Mamba and reveal opportunities for improvement.

Lucas Monteiro Paes, Nivedha Sivakumar, Yinong Oliver Wang et al.

Generative models are often deployed to make decisions on behalf of users, such as vision-language models (VLMs) identifying which person in a room is a doctor to help visually impaired individuals. Yet, VLM decisions are influenced by the perceived demographic attributes of people in the input, which can lead to biased outcomes like failing to identify women as doctors. Moreover, when reducing bias leads to performance loss, users may have varying needs for balancing bias mitigation with overall model capabilities, highlighting the demand for methods that enable controllable bias reduction during inference. Activation steering is a popular approach for inference-time controllability that has shown potential in inducing safer behavior in large language models (LLMs). However, we observe that current steering methods struggle to correct biases, where equiprobable outcomes across demographic groups are required. To address this, we propose Direct Steering Optimization (DSO) which uses reinforcement learning to find linear transformations for steering activations, tailored to mitigate bias while maintaining control over model performance. We demonstrate that DSO achieves state-of-the-art trade-off between fairness and capabilities on both VLMs and LLMs, while offering practitioners inference-time control over the trade-off. Overall, our work highlights the benefit of designing steering strategies that are directly optimized to control model behavior, providing more effective bias intervention than methods that rely on pre-defined heuristics for controllability.

Falaah Arif Khan, Nivedha Sivakumar, Yinong Oliver Wang et al.

Large language models (LLMs) have achieved impressive performance, leading to their widespread adoption as decision-support tools in resource-constrained contexts like hiring and admissions. There is, however, scientific consensus that AI systems can reflect and exacerbate societal biases, raising concerns about identity-based harm when used in critical social contexts. Prior work has laid a solid foundation for assessing bias in LLMs by evaluating demographic disparities in different language reasoning tasks. In this work, we extend single-axis fairness evaluations to examine intersectional bias, recognizing that when multiple axes of discrimination intersect, they create distinct patterns of disadvantage. We create a new benchmark called WinoIdentity by augmenting the WinoBias dataset with 25 demographic markers across 10 attributes, including age, nationality, and race, intersected with binary gender, yielding 245,700 prompts to evaluate 50 distinct bias patterns. Focusing on harms of omission due to underrepresentation, we investigate bias through the lens of uncertainty and propose a group (un)fairness metric called Coreference Confidence Disparity which measures whether models are more or less confident for some intersectional identities than others. We evaluate five recently published LLMs and find confidence disparities as high as 40% along various demographic attributes including body type, sexual orientation and socio-economic status, with models being most uncertain about doubly-disadvantaged identities in anti-stereotypical settings. Surprisingly, coreference confidence decreases even for hegemonic or privileged markers, indicating that the recent impressive performance of LLMs is more likely due to memorization than logical reasoning. Notably, these are two independent failures in value alignment and validity that can compound to cause social harm.

Krishna Patel, Nivedha Sivakumar, Barry-John Theobald et al.

We investigate fairness dynamics during Large Language Model (LLM) training to enable the diagnoses of biases and mitigations through training interventions like early stopping; we find that biases can emerge suddenly and do not always follow common performance metrics. We introduce two new metrics to evaluate fairness dynamics holistically during model pre-training: Average Rank and Jensen-Shannon Divergence by Parts. These metrics provide insights into the Pythia models' progression of biases in gender prediction of occupations on the WinoBias dataset. By monitoring these dynamics, we find that (1) Pythia-6.9b is biased towards men; it becomes more performant and confident predicting "male" than "female" during training, (2) via early-stopping, Pythia-6.9b can exchange 1.7% accuracy on LAMBADA for a 92.5% increase in fairness, and (3) larger models can exhibit more bias; Pythia-6.9b makes more assumptions about gender than Pythia-160m, even when a subject's gender is not specified.

Pau Rodriguez, Michal Klein, Eleonora Gualdoni et al.

The growing use of generative models in daily life calls for efficient mechanisms to control their generation, to e.g., produce safe content or provide users with tools to explore style changes. Ideally, such mechanisms should require low volume of unpaired data (i.e., without explicit preference), and should be cheap, both at train and inference time, while preserving output quality. Recent research has shown that such mechanisms can be obtained by intervening exclusively on model activations, with the goal of correcting distributional differences between activations seen when using prompts from a source vs. a target set (e.g., toxic and non-toxic sentences). While cheap, these fast methods are inherently crude: their maps are tuned locally, not accounting for their impact on downstream layers, resulting in interventions that cause unintended shifts when used out-of-sample. We propose in this work linear end-to-end activation steering (LinEAS), an approach trained with a global loss that accounts simultaneously for all layer-wise distributional shifts. In addition to being more robust, the loss used to train LinEAS can be regularized with sparsifying norms, which can automatically carry out neuron selection. LinEAS only requires a handful of unpaired samples to be effective, and beats similar baselines on toxicity mitigation in language models, becoming competitive with oracle-dependent methods that have access to strong supervision. LinEAS is modality-agnostic and we empirically find that it outperforms existing activation steering methods at mitigating and including new concepts at the output of single-step text-to-image generation models.

Arno Blaas, Adam Goliński, Andrew Miller et al. · apple-ml

We consider robustness to distribution shifts in the context of diagnostic models in healthcare, where the prediction target $Y$, e.g., the presence of a disease, is causally upstream of the observations $X$, e.g., a biomarker. Distribution shifts may occur, for instance, when the training data is collected in a domain with patients having particular demographic characteristics while the model is deployed on patients from a different demographic group. In the domain of applied ML for health, it is common to predict $Y$ from $X$ without considering further information about the patient. However, beyond the direct influence of the disease $Y$ on biomarker $X$, a predictive model may learn to exploit confounding dependencies (or shortcuts) between $X$ and $Y$ that are unstable under certain distribution shifts. In this work, we highlight a data generating mechanism common to healthcare settings and discuss how recent theoretical results from the causality literature can be applied to build robust predictive models. We theoretically show why ignoring covariates as well as common invariant learning approaches will in general not yield robust predictors in the studied setting, while including certain covariates into the prediction model will. In an extensive simulation study, we showcase the robustness (or lack thereof) of different predictors under various data generating processes. Lastly, we analyze the performance of the different approaches using the PTB-XL dataset, a public dataset of annotated ECG recordings.

Aparna R. Joshi, Xavier Suau, Nivedha Sivakumar et al.

As the use of deep learning in high impact domains becomes ubiquitous, it is increasingly important to assess the resilience of models. One such high impact domain is that of face recognition, with real world applications involving images affected by various degradations, such as motion blur or high exposure. Moreover, images captured across different attributes, such as gender and race, can also challenge the robustness of a face recognition algorithm. While traditional summary statistics suggest that the aggregate performance of face recognition models has continued to improve, these metrics do not directly measure the robustness or fairness of the models. Visual Psychophysics Sensitivity Analysis (VPSA) [1] provides a way to pinpoint the individual causes of failure by way of introducing incremental perturbations in the data. However, perturbations may affect subgroups differently. In this paper, we propose a new fairness evaluation based on robustness in the form of a generic framework that extends VPSA. With this framework, we can analyze the ability of a model to perform fairly for different subgroups of a population affected by perturbations, and pinpoint the exact failure modes for a subgroup by measuring targeted robustness. With the increasing focus on the fairness of models, we use face recognition as an example application of our framework and propose to compactly visualize the fairness analysis of a model via AUC matrices. We analyze the performance of common face recognition models and empirically show that certain subgroups are at a disadvantage when images are perturbed, thereby uncovering trends that were not visible using the model's performance on subgroups without perturbations.

Arno Blaas, Xavier Suau, Jason Ramapuram et al.

Image augmentations applied during training are crucial for the generalization performance of image classifiers. Therefore, a large body of research has focused on finding the optimal augmentation policy for a given task. Yet, RandAugment [2], a simple random augmentation policy, has recently been shown to outperform existing sophisticated policies. Only Adversarial AutoAugment (AdvAA) [11], an approach based on the idea of adversarial training, has shown to be better than RandAugment. In this paper, we show that random augmentations are still competitive compared to an optimal adversarial approach, as well as to simple curricula, and conjecture that the success of AdvAA is due to the stochasticity of the policy controller network, which introduces a mild form of curriculum.

Xavier Suau, Luca Zappella, Nicholas Apostoloff

In this paper we aim to investigate the mechanisms that guide text generation with pre-trained Transformer-based Language Models (TLMs). Grounded on the Product of Experts formulation by Hinton (1999), we describe a generative mechanism that exploits expert units which naturally exist in TLMs. Such units are responsible for detecting concepts in the input and conditioning text generation on such concepts. We describe how to identify expert units and how to activate them during inference in order to induce any desired concept in the generated output. We find that the activation of a surprisingly small amount of units is sufficient to steer text generation (as little as 3 units in a model with 345M parameters). While the objective of this work is to learn more about how TLMs work, we show that our method is effective for conditioning without fine-tuning or using extra parameters, even on fine-grained homograph concepts. Additionally, we show that our method can be used to correct gender bias present in the output of TLMs and achieves gender parity for all evaluated contexts. We compare our method with FUDGE and PPLM-BoW, and show that our approach is able to achieve gender parity at a lower perplexity. The proposed method is accessible to a wide audience thanks to its simplicity and minimal compute needs. The findings in this paper are a step forward in understanding the generative mechanisms of TLMs.

Xavier Suau, Luca Zappella, Nicholas Apostoloff

In this work we study the presence of expert units in pre-trained Transformer Models (TM), and how they impact a model's performance. We define expert units to be neurons that are able to classify a concept with a given average precision, where a concept is represented by a binary set of sentences containing the concept (or not). Leveraging the OneSec dataset (Scarlini et al., 2019), we compile a dataset of 1641 concepts that allows diverse expert units in TM to be discovered. We show that expert units are important in several ways: (1) The presence of expert units is correlated ($r^2=0.833$) with the generalization power of TM, which allows ranking TM without requiring fine-tuning on suites of downstream tasks. We further propose an empirical method to decide how accurate such experts should be to evaluate generalization. (2) The overlap of top experts between concepts provides a sensible way to quantify concept co-learning, which can be used for explainability of unknown concepts. (3) We show how to self-condition off-the-shelf pre-trained language models to generate text with a given concept by forcing the top experts to be active, without requiring re-training the model or using additional parameters.

Xavier Suau, Luca Zappella, Nicholas Apostoloff

In this paper we introduce Principal Filter Analysis (PFA), an easy to use and effective method for neural network compression. PFA exploits the correlation between filter responses within network layers to recommend a smaller network that maintain as much as possible the accuracy of the full model. We propose two algorithms: the first allows users to target compression to specific network property, such as number of trainable variable (footprint), and produces a compressed model that satisfies the requested property while preserving the maximum amount of spectral energy in the responses of each layer, while the second is a parameter-free heuristic that selects the compression used at each layer by trying to mimic an ideal set of uncorrelated responses. Since PFA compresses networks based on the correlation of their responses we show in our experiments that it gains the additional flexibility of adapting each architecture to a specific domain while compressing. PFA is evaluated against several architectures and datasets, and shows considerable compression rates without compromising accuracy, e.g., for VGG-16 on CIFAR-10, CIFAR-100 and ImageNet, PFA achieves a compression rate of 8x, 3x, and 1.4x with an accuracy gain of 0.4%, 1.4% points, and 2.4% respectively. Our tests show that PFA is competitive with state-of-the-art approaches while removing adoption barriers thanks to its practical implementation, intuitive philosophy and ease of use.