Subset selection tasks, such as anomaly detection and compound selection in AI-assisted drug discovery, are crucial for a wide range of applications. Learning subset-valued functions with neural networks has achieved great success by incorporating permutation invariance symmetry into the architecture. However, existing neural set architectures often struggle to either capture comprehensive information from the superset or address complex interactions within the input. Additionally, they often fail to perform in scenarios where superset sizes surpass available memory capacity. To address these challenges, we introduce the novel concept of the Identity Property, which requires models to integrate information from the originating set, resulting in the development of neural networks that excel at performing effective subset selection from large supersets. Moreover, we present the Hierarchical Representation of Neural Subset Selection (HORSE), an attention-based method that learns complex interactions and retains information from both the input set and the optimal subset supervision signal. Specifically, HORSE enables the partitioning of the input ground set into manageable chunks that can be processed independently and then aggregated, ensuring consistent outcomes across different partitions. Through extensive experimentation, we demonstrate that HORSE significantly enhances neural subset selection performance by capturing more complex information and surpasses state-of-the-art methods in handling large-scale inputs by a margin of up to 20%.

Large language models are usually fine-tuned to align with human preferences. However, fine-tuning a large language model can be challenging. In this work, we introduce $\textit{weak-to-strong search}$, framing the alignment of a large language model as a test-time greedy search to maximize the log-probability difference between small tuned and untuned models while sampling from the frozen large model. This method serves both as (1) a compute-efficient model up-scaling strategy that avoids directly tuning the large model and as (2) an instance of weak-to-strong generalization that enhances a strong model with weak test-time guidance.Empirically, we demonstrate the flexibility of weak-to-strong search across different tasks. In controlled-sentiment generation and summarization, we use tuned and untuned $\texttt{gpt2}$s to improve the alignment of large models without additional training.

We present a novel set of rigorous and computationally efficient topology-based complexity notions that exhibit a strong correlation with the generalization gap in modern deep neural networks (DNNs). DNNs show remarkable generalization properties, yet the source of these capabilities remains elusive, defying the established statistical learning theory. Recent studies have revealed that properties of training trajectories can be indicative of generalization. Building on this insight, state-of-the-art methods have leveraged the topology of these trajectories, particularly their fractal dimension, to quantify generalization. Most existing works compute this quantity by assuming continuous-or infinite-time training dynamics, complicating the development of practical estimators capable of accurately predicting generalization without access to test data.

We study a typical optimization model where the optimization variable is composed of multiple probability distributions. Though the model appears frequently in practice, such as for policy problems, it lacks specific analysis in the general setting. For this optimization problem, we propose a new structural condition/landscape description named generalized quasar-convexity (GQC) beyond the realms of convexity. In contrast to original quasar-convexity \citep{hinder2020near}, GQC allows an individual quasar-convex parameter $\gamma_i$ for each variable block $i$ and the smaller of $\gamma_i$ implies less block-convexity. To minimize the objective function, we consider a generalized oracle termed as the internal function that includes the standard gradient oracle as a special case.

State-of-the-art results in large language models (LLMs) often rely on scale, whichbecomes computationally expensive. This has sparked a research agenda to reducethese models' parameter counts and computational costs without significantlyimpacting their performance. Our study focuses on transformer-based LLMs,specifically targeting the computationally intensive feedforward networks (FFNs),which are less studied than attention blocks. We consider three structured linearparameterizations of the FFN using efficient low-rank and block-diagonal matrices.In contrast to many previous works that examined these approximations, our studyi) explores these structures from a training-from-scratch perspective, ii) scales upto 1.3B parameters, and iii) is conducted within recent Transformer-based LLMsrather than convolutional architectures. We demonstrate that these structures canlead to actual computational gains in various scenarios, including online decodingwhen using a pre-merge technique. Additionally, we propose a novel trainingregime, called self-guided training, aimed at improving the poor training dynamicsthat these approximations exhibit when used from initialization. Interestingly,the scaling performance of structured matrices is explored, revealing steepercurves in scaling training FLOPs, along with a favorable scaling trend in theovertraining regime. Specifically, we show that wide and structured networkscan utilize training FLOPs more efficiently, with fewer parameters and lowerloss than dense models at their optimal trade-off.

This work addresses the fundamental linear inverse problem in compressive sensing (CS) by introducing a new type of regularizing generative prior. Our proposed method utilizes ideas from classical dictionary-based CS and, in particular, sparse Bayesian learning (SBL), to integrate a strong regularization towards sparse solutions. At the same time, by leveraging the notion of conditional Gaussianity, it also incorporates the adaptability from generative models to training data. However, unlike most state-of-the-art generative models, it is able to learn from a few compressed and noisy data samples and requires no optimization algorithm for solving the inverse problem.

In this paper, we propose an innovative solution by constructing infinite tree-structured PGMs that correspond exactly to neural networks. Our research reveals that DNNs, during forward propagation, indeed perform approximations of PGM inference that are precise in this alternative PGM structure. Not only does our research complement existing studies that describe neural networks as kernel machines or infinite-sized Gaussian processes, it also elucidates a more direct approximation that DNNs make to exact inference in PGMs. Potential benefits include improved pedagogy and interpretation of DNNs, and algorithms that can merge the strengths of PGMs and DNNs.

The efficacy of large language models (LLMs) on downstream tasks usually hinges on instruction tuning, which relies critically on the quality of training data. Unfortunately, collecting high-quality and diverse data is both expensive and time-consuming. To mitigate this issue, we propose a novel Star-Agents framework, which automates the enhancement of data quality across datasets through multi-agent collaboration and assessment. The framework adopts a three-pronged strategy. It initially generates diverse instruction data with multiple LLM agents through a bespoke sampling method. Subsequently, the generated data undergo a rigorous evaluation using a dual-model method that assesses both difficulty and quality.

Deep learning-based methods significantly advance the exploration of associations among triple-wise biological entities (e.g., drug-target protein-adverse reaction), thereby facilitating drug discovery and safeguarding human health. However, existing researches only focus on entity-centric information mapping and aggregation, neglecting the crucial role of potential association patterns among different entities. To address the above limitation, we propose a novel association pattern-aware fusion method for biological entity relationship prediction, which effectively integrates the related association pattern information into entity representation learning. Additionally, to enhance the missing information of the low-order message passing, we devise a bind-relation module that considers the strong bind of low-order entity associations. Extensive experiments conducted on three biological datasets quantitatively demonstrate that the proposed method achieves about 4%-23% hit@1 improvements compared with state-of-the-art baselines. Furthermore, the interpretability of association patterns is elucidated in detail, thus revealing the intrinsic biological mechanisms and promoting it to be deployed in real-world scenarios.

The effect of regularizers such as weight decay when training deep neural networks is not well understood. We study the influence of weight decay as well as $L2$-regularization when training neural network models in which parameter matrices interact multiplicatively. This combination is of particular interest as this parametrization is common in attention layers, the workhorse of transformers. Here, key-query, as well as value-projection parameter matrices, are multiplied directly with each other: $W_K^TW_Q$ and $PW_V$. We extend previous results and show on one hand that any local minimum of a $L2$-regularized loss of the form $L(AB^\top) + \lambda (\|A\|^2 + \|B\|^2)$ coincides with a minimum of the nuclear norm-regularized loss $L(AB^\top) + \lambda\|AB^\top\|_*$, and on the other hand that the 2 losses become identical exponentially quickly during training. We thus complement existing works linking $L2$-regularization with low-rank regularization, and in particular, explain why such regularization on the matrix product affects early stages of training.Based on these theoretical insights, we verify empirically that the key-query and value-projection matrix products $W_K^TW_Q, PW_V$ within attention layers, when optimized with weight decay, as usually done in vision tasks and language modelling, indeed induce a significant reduction in the rank of $W_K^TW_Q$ and $PW_V$, even in fully online training.We find that, in accordance with existing work, inducing low rank in attention matrix products can damage language model performance, and observe advantages when decoupling weight decay in attention layers from the rest of the parameters.