As such, unimodal distributions cannot fully capture statistics of convolutional activations, which will limit performance of deep CNNs.

We study the theoretical properties of the interactive learning protocol Discriminative Feature Feedback (DFF) (Dasgupta et al., 2018). The DFF learning protocol uses feedback in the form of discriminative feature explanations. We provide the first systematic study of DFF in a general framework that is comparable to that of classical protocols such as supervised learning and online learning. We study the optimal mistake bound of DFF in the realizable and the non-realizable settings, and obtain novel structural results, as well as insights into the differences between Online Learning and settings with richer feedback such as DFF. We characterize the mistake bound in the realizable setting using a new notion of dimension. In the non-realizable setting, we provide a mistake upper bound and show that it cannot be improved in general. Our results show that unlike Online Learning, in DFF the realizable dimension is insufficient to characterize the optimal non-realizable mistake bound or the existence of no-regret algorithms.

Modern retrieval systems do not rely on a single ranking model to construct their rankings. Instead, they generally take a cascading approach where a sequence of ranking models are applied in multiple re-ranking stages. Thereby, they balance the quality of the top-K ranking with computational costs by limiting the number of documents each model re-ranks. However, the cascading approach is not the only way models can interact to form a retrieval system. We propose the concept of compound retrieval systems as a broader class of retrieval systems that apply multiple prediction models. This encapsulates cascading models but also allows other types of interactions than top-K re-ranking. In particular, we enable interactions with large language models (LLMs) which can provide relative relevance comparisons. We focus on the optimization of compound retrieval system design which uniquely involves learning where to apply the component models and how to aggregate their predictions into a final ranking. This work shows how our compound approach can combine the classic BM25 retrieval model with state-of-the-art (pairwise) LLM relevance predictions, while optimizing a given ranking metric and efficiency target. Our experimental results show optimized compound retrieval systems provide better trade-offs between effectiveness and efficiency than cascading approaches, even when applied in a self-supervised manner. With the introduction of compound retrieval systems, we hope to inspire the information retrieval field to more out-of-the-box thinking on how prediction models can interact to form rankings.

Training data compositions for Large Language Models (LLMs) can significantly affect their downstream performance. However, a thorough data ablation study exploring large sets of candidate data mixtures is typically prohibitively expensive since the full effect is seen only after training the models; this can lead practitioners to settle for sub-optimal data mixtures. We propose an efficient method for approximating data ablations which trains individual models on subsets of a training corpus and reuses them across evaluations of combinations of subsets. In continued pre-training experiments, we find that, given an arbitrary evaluation set, the perplexity score of a single model trained on a candidate set of data is strongly correlated with perplexity scores of parameter averages of models trained on distinct partitions of that data. From this finding, we posit that researchers and practitioners can conduct inexpensive simulations of data ablations by maintaining a pool of models that were each trained on partitions of a large training corpus, and assessing candidate data mixtures by evaluating parameter averages of combinations of these models. This approach allows for substantial improvements in amortized training efficiency -- scaling only linearly with respect to new data -- by enabling reuse of previous training computation, opening new avenues for improving model performance through rigorous, incremental data assessment and mixing.

The democratization of machine learning systems has made the process of fine-tuning accessible to a large number of practitioners, leading to a wide range of open-source models fine-tuned on specialized tasks and datasets. Recent work has proposed to merge such models to combine their functionalities. However, prior approaches are restricted to models that are fine-tuned from the same base model. Furthermore, the final merged model is typically restricted to be of the same size as the original models. In this work, we propose a new two-step algorithm to merge models-termed PLeaS-which relaxes these constraints. First, leveraging the Permutation symmetries inherent in the two models, PLeaS partially matches nodes in each layer by maximizing alignment. Next, PLeaS computes the weights of the merged model as a layer-wise Least Squares solution to minimize the approximation error between the features of the merged model and the permuted features of the original models. into a single model of a desired size, even when the two original models are fine-tuned from different base models. We also present a variant of our method which can merge models without using data from the fine-tuning domains. We demonstrate our method to merge ResNet models trained with shared and different label spaces, and show that we can perform better than the state-of-the-art merging methods by 8 to 15 percentage points for the same target compute while merging models trained on DomainNet and on fine-grained classification tasks.

Many real-world systems can be represented as sets of interacting components. Examples of such systems include computational systems such as query processors, natural systems such as cells, and social systems such as families. Many approaches have been proposed in traditional (associational) machine learning to model such structured systems, including statistical relational models and graph neural networks. Despite this prior work, existing approaches to estimating causal effects typically treat such systems as single units, represent them with a fixed set of variables and assume a homogeneous data-generating process. We study a compositional approach for estimating individual treatment effects (ITE) in structured systems, where each unit is represented by the composition of multiple heterogeneous components. This approach uses a modular architecture to model potential outcomes at each component and aggregates component-level potential outcomes to obtain the unit-level potential outcomes. We discover novel benefits of the compositional approach in causal inference - systematic generalization to estimate counterfactual outcomes of unseen combinations of components and improved overlap guarantees between treatment and control groups compared to the classical methods for causal effect estimation. We also introduce a set of novel environments for empirically evaluating the compositional approach and demonstrate the effectiveness of our approach using both simulated and real-world data.

While Transformers underpin modern large language models (LMs), there is a growing list of alternative architectures with new capabilities, promises, and tradeoffs. This makes choosing the right LM architecture challenging. Recently-proposed $\textit{hybrid architectures}$ seek a best-of-all-worlds approach that reaps the benefits of all architectures. Hybrid design is difficult for two reasons: it requires manual expert-driven search, and new hybrids must be trained from scratch. We propose $\textbf{Manticore}$, a framework that addresses these challenges. Manticore $\textit{automates the design of hybrid architectures}$ while reusing pretrained models to create $\textit{pretrained}$ hybrids. Our approach augments ideas from differentiable Neural Architecture Search (NAS) by incorporating simple projectors that translate features between pretrained blocks from different architectures. We then fine-tune hybrids that combine pretrained models from different architecture families -- such as the GPT series and Mamba -- end-to-end. With Manticore, we enable LM selection without training multiple models, the construction of pretrained hybrids from existing pretrained models, and the ability to $\textit{program}$ pretrained hybrids to have certain capabilities. Manticore hybrids outperform existing manually-designed hybrids, achieve strong performance on Long Range Arena (LRA) tasks, and can improve on pretrained transformers and state space models.

In many applications, model ensembling proves to be better than a single predictive model. Hence, it is the most common post-processing technique in Automated Machine Learning (AutoML). The most popular frameworks use ensembles at the expense of reducing the interpretability of the final models. In our work, we propose cattleia - an application that deciphers the ensembles for regression, multiclass, and binary classification tasks. This tool works with models built by three AutoML packages: auto-sklearn, AutoGluon, and FLAML. The given ensemble is analyzed from different perspectives. We conduct a predictive performance investigation through evaluation metrics of the ensemble and its component models. We extend the validation perspective by introducing new measures to assess the diversity and complementarity of the model predictions. Moreover, we apply explainable artificial intelligence (XAI) techniques to examine the importance of variables. Summarizing obtained insights, we can investigate and adjust the weights with a modification tool to tune the ensemble in the desired way. The application provides the aforementioned aspects through dedicated interactive visualizations, making it accessible to a diverse audience. We believe the cattleia can support users in decision-making and deepen the comprehension of AutoML frameworks.

In longitudinal observational studies with a time-to-event outcome, a common objective in causal analysis is to estimate the causal survival curve under hypothetical intervention scenarios within the study cohort. The g-formula is a particularly useful tool for this analysis. To enhance the traditional parametric g-formula approach, we developed a more adaptable Bayesian g-formula estimator. This estimator facilitates both longitudinal predictive and causal inference. It incorporates Bayesian additive regression trees in the modeling of the time-evolving generative components, aiming to mitigate bias due to model misspecification. Specifically, we introduce a more general class of g-formulas for discrete survival data. These formulas can incorporate the longitudinal balancing scores, which serve as an effective method for dimension reduction and are vital when dealing with an expanding array of time-varying confounders. The minimum sufficient formulation of these longitudinal balancing scores is linked to the nature of treatment regimes, whether static or dynamic. For each type of treatment regime, we provide posterior sampling algorithms, which are grounded in the Bayesian additive regression trees framework. We have conducted simulation studies to illustrate the empirical performance of our proposed Bayesian g-formula estimators, and to compare them with existing parametric estimators. We further demonstrate the practical utility of our methods in real-world scenarios using data from the Yale New Haven Health System's electronic health records.

Current adversarial attack algorithms, where an adversary changes a text to fool a victim model, have been repeatedly shown to be effective against text classifiers. These attacks, however, generally assume that the victim model is monolingual and cannot be used to target multilingual victim models, a significant limitation given the increased use of these models. For this reason, in this work we propose an approach to fine-tune a multilingual paraphrase model with an adversarial objective so that it becomes able to generate effective adversarial examples against multilingual classifiers. The training objective incorporates a set of pre-trained models to ensure text quality and language consistency of the generated text. In addition, all the models are suitably connected to the generator by vocabulary-mapping matrices, allowing for full end-to-end differentiability of the overall training pipeline. The experimental validation over two multilingual datasets and five languages has shown the effectiveness of the proposed approach compared to existing baselines, particularly in terms of query efficiency. We also provide a detailed analysis of the generated attacks and discuss limitations and opportunities for future research.