Rahul Nair

Svetoslav Nizhnichenkov, Rahul Nair, Elizabeth Daly et al.

We investigate how successful bias mitigation reshapes the embedding space of encoder-only and decoder-only foundation models, offering an internal audit of model behaviour through representational analysis. Using BERT and Llama2 as representative architectures, we assess the shifts in associations between gender and occupation terms by comparing baseline and bias-mitigated variants of the models. Our findings show that bias mitigation reduces gender-occupation disparities in the embedding space, leading to more neutral and balanced internal representations. These representational shifts are consistent across both model types, suggesting that fairness improvements can manifest as interpretable and geometric transformations. These results position embedding analysis as a valuable tool for understanding and validating the effectiveness of debiasing methods in foundation models. To further promote the assessment of decoder-only models, we introduce WinoDec, a dataset consisting of 4,000 sequences with gender and occupation terms, and release it to the general public. (https://github.com/winodec/wino-dec)

Dennis Wei, Rahul Nair, Amit Dhurandhar et al.

Interpretable and explainable machine learning has seen a recent surge of interest. We focus on safety as a key motivation behind the surge and make the relationship between interpretability and safety more quantitative. Toward assessing safety, we introduce the concept of maximum deviation via an optimization problem to find the largest deviation of a supervised learning model from a reference model regarded as safe. We then show how interpretability facilitates this safety assessment. For models including decision trees, generalized linear and additive models, the maximum deviation can be computed exactly and efficiently. For tree ensembles, which are not regarded as interpretable, discrete optimization techniques can still provide informative bounds. For a broader class of piecewise Lipschitz functions, we leverage the multi-armed bandit literature to show that interpretability produces tighter (regret) bounds on the maximum deviation. We present case studies, including one on mortgage approval, to illustrate our methods and the insights about models that may be obtained from deviation maximization.

Daniel Karl I. Weidele, Shazia Afzal, Abel N. Valente et al.

We present AutoDOViz, an interactive user interface for automated decision optimization (AutoDO) using reinforcement learning (RL). Decision optimization (DO) has classically being practiced by dedicated DO researchers where experts need to spend long periods of time fine tuning a solution through trial-and-error. AutoML pipeline search has sought to make it easier for a data scientist to find the best machine learning pipeline by leveraging automation to search and tune the solution. More recently, these advances have been applied to the domain of AutoDO, with a similar goal to find the best reinforcement learning pipeline through algorithm selection and parameter tuning. However, Decision Optimization requires significantly more complex problem specification when compared to an ML problem. AutoDOViz seeks to lower the barrier of entry for data scientists in problem specification for reinforcement learning problems, leverage the benefits of AutoDO algorithms for RL pipeline search and finally, create visualizations and policy insights in order to facilitate the typical interactive nature when communicating problem formulation and solution proposals between DO experts and domain experts. In this paper, we report our findings from semi-structured expert interviews with DO practitioners as well as business consultants, leading to design requirements for human-centered automation for DO with RL. We evaluate a system implementation with data scientists and find that they are significantly more open to engage in DO after using our proposed solution. AutoDOViz further increases trust in RL agent models and makes the automated training and evaluation process more comprehensible. As shown for other automation in ML tasks, we also conclude automation of RL for DO can benefit from user and vice-versa when the interface promotes human-in-the-loop.

Jasmina Gajcin, James McCarthy, Rahul Nair et al.

A well-defined reward function is crucial for successful training of an reinforcement learning (RL) agent. However, defining a suitable reward function is a notoriously challenging task, especially in complex, multi-objective environments. Developers often have to resort to starting with an initial, potentially misspecified reward function, and iteratively adjusting its parameters, based on observed learned behavior. In this work, we aim to automate this process by proposing ITERS, an iterative reward shaping approach using human feedback for mitigating the effects of a misspecified reward function. Our approach allows the user to provide trajectory-level feedback on agent's behavior during training, which can be integrated as a reward shaping signal in the following training iteration. We also allow the user to provide explanations of their feedback, which are used to augment the feedback and reduce user effort and feedback frequency. We evaluate ITERS in three environments and show that it can successfully correct misspecified reward functions.

Elizabeth M. Daly, Massimiliano Mattetti, Öznur Alkan et al.

AI solutions are heavily dependant on the quality and accuracy of the input training data, however the training data may not always fully reflect the most up-to-date policy landscape or may be missing business logic. The advances in explainability have opened the possibility of allowing users to interact with interpretable explanations of ML predictions in order to inject modifications or constraints that more accurately reflect current realities of the system. In this paper, we present a solution which leverages the predictive power of ML models while allowing the user to specify modifications to decision boundaries. Our interactive overlay approach achieves this goal without requiring model retraining, making it appropriate for systems that need to apply instant changes to their decision making. We demonstrate that user feedback rules can be layered with the ML predictions to provide immediate changes which in turn supports learning with less data.

James McCarthy, Rahul Nair, Elizabeth Daly et al.

Explainability of Reinforcement Learning (RL) policies remains a challenging research problem, particularly when considering RL in a safety context. Understanding the decisions and intentions of an RL policy offer avenues to incorporate safety into the policy by limiting undesirable actions. We propose the use of a Boolean Decision Rules model to create a post-hoc rule-based summary of an agent's policy. We evaluate our proposed approach using a DQN agent trained on an implementation of a lava gridworld and show that, given a hand-crafted feature representation of this gridworld, simple generalised rules can be created, giving a post-hoc explainable summary of the agent's policy. We discuss possible avenues to introduce safety into a RL agent's policy by using rules generated by this rule-based model as constraints imposed on the agent's policy, as well as discuss how creating simple rule summaries of an agent's policy may help in the debugging process of RL agents.

Amit Dhurandhar, Vijil Chenthamarakshan, Dennis Wei et al.

Transformers are arguably the preferred architecture for language generation. In this paper, inspired by continued fractions, we introduce a new function class for generative modeling. The architecture family implementing this function class is named CoFrGeNets - Continued Fraction Generative Networks. We design novel architectural components based on this function class that can replace Multi-head Attention and Feed-Forward Networks in Transformer blocks while requiring much fewer parameters. We derive custom gradient formulations to optimize the proposed components more accurately and efficiently than using standard PyTorch-based gradients. Our components are a plug-in replacement requiring little change in training or inference procedures that have already been put in place for Transformer-based models thus making our approach easy to incorporate in large industrial workflows. We experiment on two very different transformer architectures GPT2-xl (1.5B) and Llama3 (3.2B), where the former we pre-train on OpenWebText and GneissWeb, while the latter we pre-train on the docling data mix which consists of nine different datasets. Results show that the performance on downstream classification, Q\& A, reasoning and text understanding tasks of our models is competitive and sometimes even superior to the original models with $\frac{2}{3}$ to $\frac{1}{2}$ the parameters and shorter pre-training time. We believe that future implementations customized to hardware will further bring out the true potential of our architectures.

Swagatam Haldar, Diptikalyan Saha, Dennis Wei et al.

Understanding the differences between machine learning (ML) models is of interest in scenarios ranging from choosing amongst a set of competing models, to updating a deployed model with new training data. In these cases, we wish to go beyond differences in overall metrics such as accuracy to identify where in the feature space do the differences occur. We formalize this problem of model differencing as one of predicting a dissimilarity function of two ML models' outputs, subject to the representation of the differences being human-interpretable. Our solution is to learn a Joint Surrogate Tree (JST), which is composed of two conjoined decision tree surrogates for the two models. A JST provides an intuitive representation of differences and places the changes in the context of the models' decision logic. Context is important as it helps users to map differences to an underlying mental model of an AI system. We also propose a refinement procedure to increase the precision of a JST. We demonstrate, through an empirical evaluation, that such contextual differencing is concise and can be achieved with no loss in fidelity over naive approaches.

Rahul Nair

We consider an aggregated human-AI collaboration aimed at generating a joint interpretable model. The model takes the form of Boolean decision rules, where human input is provided in the form of logical conditions or as partial templates. This focus on the combined construction of a model offers a different perspective on joint decision making. Previous efforts have typically focused on aggregating outcomes rather than decisions logic. We demonstrate the proposed approach through two examples and highlight the usefulness and challenges of the approach.

Bhanu Tokas, Rahul Nair, Hannah Kerner

Most ML datasets today contain biases. When we train models on these datasets, they often not only learn these biases but can worsen them -- a phenomenon known as bias amplification. Several co-occurrence-based metrics have been proposed to measure bias amplification in classification datasets. They measure bias amplification between a protected attribute (e.g., gender) and a task (e.g., cooking). These metrics also support fine-grained bias analysis by identifying the direction in which a model amplifies biases. However, co-occurrence-based metrics have limitations -- some fail to measure bias amplification in balanced datasets, while others fail to measure negative bias amplification. To solve these issues, recent work proposed a predictability-based metric called leakage amplification (LA). However, LA cannot identify the direction in which a model amplifies biases. We propose Directional Predictability Amplification (DPA), a predictability-based metric that is (1) directional, (2) works with balanced and unbalanced datasets, and (3) correctly identifies positive and negative bias amplification. DPA eliminates the need to evaluate models on multiple metrics to verify these three aspects. DPA also improves over prior predictability-based metrics like LA: it is less sensitive to the choice of attacker function (a hyperparameter in predictability-based metrics), reports scores within a bounded range, and accounts for dataset bias by measuring relative changes in predictability. Our experiments on well-known datasets like COMPAS (a tabular dataset), COCO, and ImSitu (image datasets) show that DPA is the most reliable metric to measure bias amplification in classification problems. To compare DPA with existing bias amplification metrics, we released a one-stop library of major bias amplification metrics at https://github.com/kerner-lab/Bias-Amplification.

Nico Wagner, Michael Desmond, Rahul Nair et al.

LLM-as-a-Judge is a widely used method for evaluating the performance of Large Language Models (LLMs) across various tasks. We address the challenge of quantifying the uncertainty of LLM-as-a-Judge evaluations. While uncertainty quantification has been well-studied in other domains, applying it effectively to LLMs poses unique challenges due to their complex decision-making capabilities and computational demands. In this paper, we introduce a novel method for quantifying uncertainty designed to enhance the trustworthiness of LLM-as-a-Judge evaluations. The method quantifies uncertainty by analyzing the relationships between generated assessments and possible ratings. By cross-evaluating these relationships and constructing a confusion matrix based on token probabilities, the method derives labels of high or low uncertainty. We evaluate our method across multiple benchmarks, demonstrating a strong correlation between the accuracy of LLM evaluations and the derived uncertainty scores. Our findings suggest that this method can significantly improve the reliability and consistency of LLM-as-a-Judge evaluations.

Amit Dhurandhar, Rahul Nair, Moninder Singh et al.

Evaluation and ranking of large language models (LLMs) has become an important problem with the proliferation of these models and their impact. Evaluation methods either require human responses which are expensive to acquire or use pairs of LLMs to evaluate each other which can be unreliable. In this paper, we provide a novel perspective where, given a dataset of prompts (viz. questions, instructions, etc.) and a set of LLMs, we rank them without access to any ground truth or reference responses. Inspired by real life where both an expert and a knowledgeable person can identify a novice our main idea is to consider triplets of models, where each one of them evaluates the other two, correctly identifying the worst model in the triplet with high probability. We also analyze our idea and provide sufficient conditions for it to succeed. Applying this idea repeatedly, we propose two methods to rank LLMs. In experiments on different generative tasks (summarization, multiple-choice, and dialog), our methods reliably recover close to true rankings without reference data. This points to a viable low-resource mechanism for practical use.

Jan-Christoph Klie, Juan Haladjian, Marc Kirchner et al.

Annotated datasets are an essential ingredient to train, evaluate, compare and productionalize supervised machine learning models. It is therefore imperative that annotations are of high quality. For their creation, good quality management and thereby reliable quality estimates are needed. Then, if quality is insufficient during the annotation process, rectifying measures can be taken to improve it. Quality estimation is often performed by having experts manually label instances as correct or incorrect. But checking all annotated instances tends to be expensive. Therefore, in practice, usually only subsets are inspected; sizes are chosen mostly without justification or regard to statistical power and more often than not, are relatively small. Basing estimates on small sample sizes, however, can lead to imprecise values for the error rate. Using unnecessarily large sample sizes costs money that could be better spent, for instance on more annotations. Therefore, we first describe in detail how to use confidence intervals for finding the minimal sample size needed to estimate the annotation error rate. Then, we propose applying acceptance sampling as an alternative to error rate estimation We show that acceptance sampling can reduce the required sample sizes up to 50% while providing the same statistical guarantees.

Ethan Li, Anders Boesen Lindbo Larsen, Chen Zhang et al. · apple-ml, cmu

We introduce two multilingual, multimodal foundation language models that power Apple Intelligence features across Apple devices and services: i a 3B-parameter on-device model optimized for Apple silicon through architectural innovations such as KV-cache sharing and 2-bit quantization-aware training; and ii a scalable server model built on a novel Parallel-Track Mixture-of-Experts PT-MoE transformer that combines track parallelism, mixture-of-experts sparse computation, and interleaved global-local attention to deliver high quality with competitive cost on Apple's Private Cloud Compute platform. Both models are trained on large-scale multilingual and multimodal datasets sourced via responsible web crawling, licensed corpora, and high-quality synthetic data, then further refined with supervised fine-tuning and reinforcement learning on a new asynchronous platform. The resulting models support several additional languages while understanding images and executing tool calls. In public benchmarks and human evaluations, both the server model and the on-device model match or surpass comparably sized open baselines. A new Swift-centric Foundation Models framework exposes guided generation, constrained tool calling, and LoRA adapter fine-tuning, allowing developers to integrate these capabilities with a few lines of code. The latest advancements in Apple Intelligence models are grounded in our Responsible AI approach with safeguards like content filtering and locale-specific evaluation, as well as our commitment to protecting our users' privacy with innovations like Private Cloud Compute.

Buse Sibel Korkmaz, Rahul Nair, Elizabeth M. Daly et al.

Current debiasing approaches often result a degradation in model capabilities such as factual accuracy and knowledge retention. Through systematic evaluation across multiple benchmarks, we demonstrate that existing debiasing methods face fundamental trade-offs, particularly in smaller models, leading to reduced truthfulness, knowledge loss, or unintelligible outputs. To address these limitations, we propose a contrastive learning framework that learns through carefully constructed positive and negative examples. Our approach introduces contrast computation and dynamic loss scaling to balance bias mitigation with faithfulness preservation. Experimental results across multiple model scales demonstrate that our method achieves substantial improvements in both toxicity reduction and faithfulness preservation. Most importantly, we show that our framework is the first to consistently improve both metrics simultaneously, avoiding the capability degradation characteristic of existing approaches. These results suggest that explicit modeling of both positive and negative examples through contrastive learning could be a promising direction for reducing the alignment tax in language model debiasing.

Jasmina Gajcin, Erik Miehling, Rahul Nair et al.

Using LLMs to evaluate text, that is, LLM-as-a-judge, is increasingly being used at scale to augment or even replace human annotations. As such, it is imperative that we understand the potential biases and risks of doing so. In this work, we propose an approach for extracting high-level concept-based global policies from LLM-as-a-Judge. Our approach consists of two algorithms: 1) CLoVE (Contrastive Local Verifiable Explanations), which generates verifiable, concept-based, contrastive local explanations and 2) GloVE (Global Verifiable Explanations), which uses iterative clustering, summarization and verification to condense local rules into a global policy. We evaluate GloVE on seven standard benchmarking datasets for content harm detection. We find that the extracted global policies are highly faithful to decisions of the LLM-as-a-Judge. Additionally, we evaluated the robustness of global policies to text perturbations and adversarial attacks. Finally, we conducted a user study to evaluate user understanding and satisfaction with global policies.

Rahul Nair, Inge Vejsbjerg, Elizabeth Daly et al.

Humble AI (Knowles et al., 2023) argues for cautiousness in AI development and deployments through scepticism (accounting for limitations of statistical learning), curiosity (accounting for unexpected outcomes), and commitment (accounting for multifaceted values beyond performance). We present a real-world case study for humble AI in the domain of algorithmic hiring. Specifically, we evaluate virtual screening algorithms in a widely used hiring platform that matches candidates to job openings. There are several challenges in misrecognition and stereotyping in such contexts that are difficult to assess through standard fairness and trust frameworks; e.g., someone with a non-traditional background is less likely to rank highly. We demonstrate technical feasibility of how humble AI principles can be translated to practice through uncertainty quantification of ranks, entropy estimates, and a user experience that highlights algorithmic unknowns. We describe preliminary discussions with focus groups made up of recruiters. Future user studies seek to evaluate whether the higher cognitive load of a humble AI system fosters a climate of trust in its outcomes.

Rahul Nair, Bhanu Tokas, Hannah Kerner

When we train models on biased datasets, they not only reproduce data biases, but can worsen them at test time - a phenomenon called bias amplification. Many of the current bias amplification metrics (e.g., BA (MALS), DPA) measure bias amplification only in classification datasets. These metrics are ineffective for image captioning datasets, as they cannot capture the language semantics of a caption. Recent work introduced Leakage in Captioning (LIC), a language-aware bias amplification metric that understands caption semantics. However, LIC has a crucial limitation: it cannot identify the source of bias amplification in captioning models. We propose Directional Bias Amplification in Captioning (DBAC), a language-aware and directional metric that can identify when captioning models amplify biases. DBAC has two more improvements over LIC: (1) it is less sensitive to sentence encoders (a hyperparameter in language-aware metrics), and (2) it provides a more accurate estimate of bias amplification in captions. Our experiments on gender and race attributes in the COCO captions dataset show that DBAC is the only reliable metric to measure bias amplification in captions.

Buse Sibel Korkmaz, Rahul Nair, Elizabeth M. Daly et al.

Foundation models require fine-tuning to ensure their generative outputs align with intended results for specific tasks. Automating this fine-tuning process is challenging, as it typically needs human feedback that can be expensive to acquire. We present AutoRefine, a method that leverages reinforcement learning for targeted fine-tuning, utilizing direct feedback from measurable performance improvements in specific downstream tasks. We demonstrate the method for a problem arising in algorithmic hiring platforms where linguistic biases influence a recommendation system. In this setting, a generative model seeks to rewrite given job specifications to receive more diverse candidate matches from a recommendation engine which matches jobs to candidates. Our model detects and regulates biases in job descriptions to meet diversity and fairness criteria. The experiments on a public hiring dataset and a real-world hiring platform showcase how large language models can assist in identifying and mitigation biases in the real world.

Rahul Nair, Gabriel Tseng, Esther Rolf et al.

Previous studies showed that image datasets lacking geographic diversity can lead to biased performance in models trained on them. While earlier work studied general-purpose image datasets (e.g., ImageNet) and simple tasks like image recognition, we investigated geo-biases in real-world driving datasets on a more complex task: instance segmentation. We examined if instance segmentation models trained on European driving scenes (Eurocentric models) are geo-biased. Consistent with previous work, we found that Eurocentric models were geo-biased. Interestingly, we found that geo-biases came from classification errors rather than localization errors, with classification errors alone contributing 10-90% of the geo-biases in segmentation and 19-88% of the geo-biases in detection. This showed that while classification is geo-biased, localization (including detection and segmentation) is geographically robust. Our findings show that in region-specific models (e.g., Eurocentric models), geo-biases from classification errors can be significantly mitigated by using coarser classes (e.g., grouping car, bus, and truck as 4-wheeler).

Öznur Alkan, Dennis Wei, Massimiliano Mattetti et al.

Machine learning models may involve decision boundaries that change over time due to updates to rules and regulations, such as in loan approvals or claims management. However, in such scenarios, it may take time for sufficient training data to accumulate in order to retrain the model to reflect the new decision boundaries. While work has been done to reinforce existing decision boundaries, very little has been done to cover these scenarios where decision boundaries of the ML models should change in order to reflect new rules. In this paper, we focus on user-provided feedback rules as a way to expedite the ML models update process, and we formally introduce the problem of pre-processing training data to edit an ML model in response to feedback rules such that once the model is retrained on the pre-processed data, its decision boundaries align more closely with the rules. To solve this problem, we propose a novel data augmentation method, the Feedback Rule-Based Oversampling Technique. Extensive experiments using different ML models and real world datasets demonstrate the effectiveness of the method, in particular the benefit of augmentation and the ability to handle many feedback rules.

Jasmina Gajcin, Rahul Nair, Tejaswini Pedapati et al.

In complex tasks where the reward function is not straightforward and consists of a set of objectives, multiple reinforcement learning (RL) policies that perform task adequately, but employ different strategies can be trained by adjusting the impact of individual objectives on reward function. Understanding the differences in strategies between policies is necessary to enable users to choose between offered policies, and can help developers understand different behaviors that emerge from various reward functions and training hyperparameters in RL systems. In this work we compare behavior of two policies trained on the same task, but with different preferences in objectives. We propose a method for distinguishing between differences in behavior that stem from different abilities from those that are a consequence of opposing preferences of two RL agents. Furthermore, we use only data on preference-based differences in order to generate contrasting explanations about agents' preferences. Finally, we test and evaluate our approach on an autonomous driving task and compare the behavior of a safety-oriented policy and one that prefers speed.

Jochen Görtler, Fred Hohman, Dominik Moritz et al.

The confusion matrix, a ubiquitous visualization for helping people evaluate machine learning models, is a tabular layout that compares predicted class labels against actual class labels over all data instances. We conduct formative research with machine learning practitioners at Apple and find that conventional confusion matrices do not support more complex data-structures found in modern-day applications, such as hierarchical and multi-output labels. To express such variations of confusion matrices, we design an algebra that models confusion matrices as probability distributions. Based on this algebra, we develop Neo, a visual analytics system that enables practitioners to flexibly author and interact with hierarchical and multi-output confusion matrices, visualize derived metrics, renormalize confusions, and share matrix specifications. Finally, we demonstrate Neo's utility with three model evaluation scenarios that help people better understand model performance and reveal hidden confusions.

Sergiy Zhuk, Jonathan P. Epperlein, Rahul Nair et al.

Intra-operative identification of malignant versus benign or healthy tissue is a major challenge in fluorescence guided cancer surgery. We propose a perfusion quantification method for computer-aided interpretation of subtle differences in dynamic perfusion patterns which can be used to distinguish between normal tissue and benign or malignant tumors intra-operatively in real-time by using multispectral endoscopic videos. The method exploits the fact that vasculature arising from cancer angiogenesis gives tumors differing perfusion patterns from the surrounding tissue, and defines a signature of tumor which could be used to differentiate tumors from normal tissues. Experimental evaluation of our method on a cohort of colorectal cancer surgery endoscopic videos suggests that the proposed tumor signature is able to successfully discriminate between healthy, cancerous and benign tissue with 95% accuracy.

Mojmir Mutny, Rahul Nair, Jens-Malte Gottfried

The Multipath effect in Time-of-Flight(ToF) cameras still remains to be a challenging problem that hinders further processing of 3D data information. Based on the evidence from previous literature, we explored the possibility of using machine learning techniques to correct this effect. Firstly, we created two new datasets of of ToF images rendered via ToF simulator of LuxRender. These two datasets contain corners in multiple orientations and with different material properties. We chose scenes with corners as multipath effects are most pronounced in corners. Secondly, we used this dataset to construct a learning model to predict real valued corrections to the ToF data using Random Forests. We found out that in our smaller dataset we were able to predict real valued correction and improve the quality of depth images significantly by removing multipath bias. With our algorithm, we improved relative per-pixel error from average value of 19% to 3%. Additionally, variance of the error was lowered by an order of magnitude.