ICML 2024 Papers

Pre-trained Transformers inherently possess the characteristic of sparse activation, where only a small fraction of the neurons are activated for each token. While sparse activation has been explored through post-training methods, its potential in pre-training remains untapped. In this work, we first study how activation properties change during pre-training. Our examination reveals that Transformers exhibit sparse activation throughout the majority of the pre-training process while the activation correlation keeps evolving as training progresses. Leveraging this observation, we propose Switchable Sparse-Dense Learning (SSD). SSD adaptively switches between the Mixtures-of-Experts (MoE) based sparse training and the conventional dense training during the pre-training process, leveraging the efficiency of sparse training and avoiding the static activation correlation of sparse training. Compared to dense training, SSD achieves comparable performance with identical model size and reduces pre-training costs. Moreover, the models trained with SSD can be directly used as MoE models for sparse inference and achieve the same performance as dense models with up to $2\times$ faster inference speed. Codes are available at https://github.com/thunlp/moefication.

By leveraging a hierarchical structure of known classes, Hierarchical Novelty Detection (HND) offers fine-grained detection results that pair detected novel samples with their closest (known) parent class in the hierarchy. Prior knowledge on the parent class provides valuable insights to better understand these novel samples. However, traditional novelty detection methods try to separate novel samples from all known classes using uncertainty or distance based metrics so they are incapable of locating the closest known parent class. Since the novel class is also part of the hierarchy, the model can more easily get confused between samples from known classes and those from novel ones. To achieve effective HND, we propose to augment the known (leaf-level) classes with a set of novel classes, each of which is associated with one parent (i.e., non-leaf) class in the original hierarchy. Such a structure allows us to perform novel fine-grained evidence allocation to differentiate known and novel classes guided by a uniquely designed loss function. Our thorough theoretical analysis shows that fine-grained evidence allocation creates an evidence margin to more precisely separate known and novel classes. Extensive experiments conducted on real-world hierarchical datasets demonstrate the proposed model outperforms the strongest baselines and achieves the best HND performance.

Error correction codes are a crucial part of the physical communication layer, ensuring the reliable transfer of data over noisy channels. The design of optimal linear block codes capable of being efficiently decoded is of major concern, especially for short block lengths. While neural decoders have recently demonstrated their advantage over classical decoding techniques, the neural design of the codes remains a challenge. In this work, we propose for the first time a unified encoder-decoder training of binary linear block codes. To this end, we adapt the coding setting to support efficient and differentiable training of the code for end-to-end optimization over the order two Galois field. We also propose a novel Transformer model in which the self-attention masking is performed in a differentiable fashion for the efficient backpropagation of the code gradient. Our results show that (i) the proposed decoder outperforms existing neural decoding on conventional codes, (ii) the suggested framework generates codes that outperform the analogous conventional codes, and (iii) the codes we developed not only excel with our decoder but also show enhanced performance with traditional decoding techniques.

Exploiting invariant relations and mitigating spurious correlation (a.k.a., shortcut) between representation and target across varied data distributions can tackle the challenging out-of-distribution (OOD) generalization problem. In personalized federated learning (PFL), heterogeneous data distribution across local clients offers the inherent prerequisites to extract the invariant features that maintain invariant relation with target. Nevertheless, personalized features are closely entangled with spurious features in PFL since they exhibit similar variability across different clients, which makes preserving personalization knowledge and eliminating shortcuts two conflicting objectives in PFL. To address the above challenge, we analyse the heterogeneous data generation on local clients through the lens of structured causal model and propose a crucial causal signature which can distinguish personalized features from spurious features with global invariant features as the anchor. Then the causal signature is quantified as an information-theoretic constraint that facilitates the shortcut-averse personalized invariant learning on each client. Theoretical analysis demonstrates our method, FedPIN, can yield a tighter bound on generalization error than the prevalent PFL approaches when train-test distribution shift exists on clients. Moreover, we provide a theoretical guarantee on the convergence rate of FedPIN in this paper. The results of extensive experiments show that our method can achieve superior OOD generalization performance compared with the state-of-the-art competitors.

Generative models for structure-based drug design (SBDD) have shown promising results in recent years. Existing works mainly focus on how to generate molecules with higher binding affinity, ignoring the feasibility prerequisites for generated 3D poses and resulting in *false positives*. We conduct thorough studies on key factors of ill-conformational problems when applying autoregressive methods and diffusion to SBDD, including mode collapse and hybrid continuous-discrete space. In this paper, we introduce MolCRAFT, the first SBDD model that operates in the continuous parameter space, together with a novel noise reduced sampling strategy. Empirical results show that our model consistently achieves superior performance in binding affinity with more stable 3D structure, demonstrating our ability to accurately model interatomic interactions. To our best knowledge, MolCRAFT is the first to achieve reference-level Vina Scores (-6.59 kcal/mol) with comparable molecular size, outperforming other strong baselines by a wide margin (-0.84 kcal/mol). Code is available at https://github.com/AlgoMole/MolCRAFT.

A common network inference problem, arising from real-world data constraints, is how to infer a dynamic network from its time-aggregated adjacency matrix and time-varying marginals (i.e., row and column sums). Prior approaches to this problem have repurposed the classic iterative proportional fitting (IPF) procedure, also known as Sinkhorn’s algorithm, with promising empirical results. However, the statistical foundation for using IPF has not been well understood: under what settings does IPF provide principled estimation of a dynamic network from its marginals, and how well does it estimate the network? In this work, we establish such a setting, by identifying a generative network model whose maximum likelihood estimates are recovered by IPF. Our model both reveals implicit assumptions on the use of IPF in such settings and enables new analyses, such as structure-dependent error bounds on IPF’s parameter estimates. When IPF fails to converge on sparse network data, we introduce a principled algorithm that guarantees IPF converges under minimal changes to the network structure. Finally, we conduct experiments with synthetic and real-world data, which demonstrate the practical value of our theoretical and algorithmic contributions.

An important aspect in developing language models that interact with humans is aligning their behavior to be useful and unharmful for their human users. This is usually achieved by tuning the model in a way that enhances desired behaviors and inhibits undesired ones, a process referred to as alignment. In this paper, we propose a theoretical approach called Behavior Expectation Bounds (BEB) which allows us to formally investigate several inherent characteristics and limitations of alignment in large language models. Importantly, we prove that within the limits of this framework, for any behavior that has a finite probability of being exhibited by the model, there exist prompts that can trigger the model into outputting this behavior, with probability that increases with the length of the prompt. This implies that any alignment process that attenuates an undesired behavior but does not remove it altogether, is not safe against adversarial prompting attacks. Furthermore, our framework hints at the mechanism by which leading alignment approaches such as reinforcement learning from human feedback make the LLM prone to being prompted into the undesired behaviors. This theoretical result is being experimentally demonstrated in large scale by the so called contemporary "chatGPT jailbreaks", where adversarial users trick the LLM into breaking its alignment guardrails by triggering it into acting as a malicious persona. Our results expose fundamental limitations in alignment of LLMs and bring to the forefront the need to devise reliable mechanisms for ensuring AI safety.

As AI model size grows, neural *scaling laws* have become a crucial tool to predict the improvements of large models when increasing capacity and the size of original (human or natural) training data. Yet, the widespread use of popular models means that the ecosystem of online data and text will co-evolve to progressively contain increased amounts of synthesized data. In this paper we ask: *How will the scaling laws change in the inevitable regime where synthetic data makes its way into the training corpus?* Will future models, still improve, or be doomed to degenerate up to total *(model) collapse*? We develop a theoretical framework of model collapse through the lens of scaling laws. We discover a wide range of decay phenomena, analyzing loss of scaling, shifted scaling with number of generations, the ''un-learning" of skills, and grokking when mixing human and synthesized data. Our theory is validated by large-scale experiments with a transformer on an arithmetic task and text generation using the large language model Llama2.

Spontaneous neural activity, crucial in memory, learning, and spatial navigation, often manifests itself as repetitive spatiotemporal patterns. Despite their importance, analyzing these patterns in large neural recordings remains challenging due to a lack of efficient and scalable detection methods. Addressing this gap, we introduce *convSeq*, an unsupervised method that employs backpropagation for optimizing spatiotemporal filters that effectively identify these neural patterns. Our method’s performance is validated on various synthetic data and real neural recordings, revealing spike sequences with unprecedented scalability and efficiency. Significantly surpassing existing methods in speed, *convSeq* sets a new standard for analyzing spontaneous neural activity, potentially advancing our understanding of information processing in neural circuits.

Text-to-Image (T2I) models have raised security concerns due to their potential to generate inappropriate or harmful images. In this paper, we propose UPAM, a novel framework that investigates the robustness of T2I models from the attack perspective. Unlike most existing attack methods that focus on deceiving textual defenses, UPAM aims to deceive both textual and visual defenses in T2I models. UPAM enables gradient-based optimization, offering greater effectiveness and efficiency than previous methods. Given that T2I models might not return results due to defense mechanisms, we introduce a Sphere-Probing Learning (SPL) scheme to support gradient optimization even when no results are returned. Additionally, we devise a Semantic-Enhancing Learning (SEL) scheme to finetune UPAM for generating target-aligned images. Our framework also ensures attack stealthiness. Extensive experiments demonstrate UPAM's effectiveness and efficiency.

We present Premier-TACO, a multitask feature representation learning approach designed to improve few-shot policy learning efficiency in sequential decision-making tasks. Premier-TACO leverages a subset of multitask offline datasets for pretraining a general feature representation, which captures critical environmental dynamics and is fine-tuned using minimal expert demonstrations. It advances the temporal action contrastive learning (TACO) objective, known for state-of-the-art results in visual control tasks, by incorporating a novel negative example sampling strategy. This strategy is crucial in significantly boosting TACO’s computational efficiency, making large-scale multitask offline pretraining feasible. Our extensive empirical evaluation in a diverse set of continuous control benchmarks including Deepmind Control Suite, MetaWorld, and LIBERO demonstrate Premier-TACO’s effective- ness in pretraining visual representations, significantly enhancing few-shot imitation learning of novel tasks.

The random feature (RF) approach is a well-established and efficient tool for scalable kernel methods, but existing literature has primarily focused on kernel ridge regression with random features (KRR-RF), which has limitations in handling heterogeneous data with heavy-tailed noises. This paper presents a generalization study of kernel quantile regression with random features (KQR-RF), which accounts for the non-smoothness of the check loss in KQR-RF by introducing a refined error decomposition and establishing a novel connection between KQR-RF and KRR-RF. Our study establishes the capacity-dependent learning rates for KQR-RF under mild conditions on the number of RFs, which are minimax optimal up to some logarithmic factors. Importantly, our theoretical results, utilizing a data-dependent sampling strategy, can be extended to cover the agnostic setting where the target quantile function may not precisely align with the assumed kernel space. By slightly modifying our assumptions, the capacity-dependent error analysis can also be applied to cases with Lipschitz continuous losses, enabling broader applications in the machine learning community. To validate our theoretical findings, simulated experiments and a real data application are conducted.

We propose MM-Vet, an evaluation benchmark that examines large multimodal models (LMMs) on complicated multimodal tasks. Recent LMMs have shown various intriguing abilities, such as solving math problems written on the blackboard, reasoning about events and celebrities in news images, and explaining visual jokes. Rapid model advancements pose challenges to evaluation benchmark development. Problems include: (1) How to systematically structure and evaluate the complicated multimodal tasks; (2) How to design evaluation metrics that work well across question and answer types; and (3) How to give model insights beyond a simple performance ranking. To this end, we present MM-Vet, designed based on the insight that the intriguing ability to solve complicated tasks is often achieved by a generalist model being able to integrate different core vision-language (VL) capabilities. MM-Vet defines 6 core VL capabilities and examines the 16 integrations of interest derived from the capability combination. For evaluation metrics, we propose an LLM-based evaluator for open-ended outputs. The evaluator enables the evaluation across different question types and answer styles, resulting in a unified scoring metric. We evaluate representative LMMs on MM-Vet, providing insights into the capabilities of different LMM system paradigms and models.

Dynamic regret minimization offers a principled way for non-stationary online learning, where the algorithm's performance is evaluated against changing comparators. Prevailing methods often employ a two-layer online ensemble, consisting of a group of base learners with different configurations and a meta learner that combines their outputs. Given the evident computational overhead associated with two-layer algorithms, this paper investigates how to attain optimal dynamic regret *without* deploying a model ensemble. To this end, we introduce the notion of *underlying dynamic regret*, a specific form of the general dynamic regret that can encompass many applications of interest. We show that almost optimal dynamic regret can be obtained using a single-layer model alone. This is achieved by an adaptive restart equipped with wavelet detection, wherein a novel streaming wavelet operator is introduced to online update the wavelet coefficients via a carefully designed binary indexed tree. We apply our method to the *online label shift* adaptation problem, leading to new algorithms with optimal dynamic regret and significantly improved computation/storage efficiency compared to prior arts. Extensive experiments validate our proposal.

By jointly learning multiple tasks, multi-task learning (MTL) can leverage the shared knowledge across tasks, resulting in improved data efficiency and generalization performance. However, a major challenge in MTL lies in the presence of conflicting gradients, which can hinder the fair optimization of some tasks and subsequently impede MTL's ability to achieve better overall performance. Inspired by fair resource allocation in communication networks, we formulate the optimization of MTL as a utility maximization problem, where the loss decreases across tasks are maximized under different fairness measurements. To address the problem, we propose FairGrad, a novel optimization objective. FairGrad not only enables flexible emphasis on certain tasks but also achieves a theoretical convergence guarantee. Extensive experiments demonstrate that our method can achieve state-of-the-art performance among gradient manipulation methods on a suite of multi-task benchmarks in supervised learning and reinforcement learning. Furthermore, we incorporate the idea of $\alpha$-fairness into the loss functions of various MTL methods. Extensive empirical studies demonstrate that their performance can be significantly enhanced. Code is available at https://github.com/OptMN-Lab/fairgrad.

GFlowNets are a promising alternative to MCMC sampling for discrete compositional random variables. Training GFlowNets requires repeated evaluations of the unnormalized target distribution, or reward function. However, for large-scale posterior sampling, this may be prohibitive since it incurs traversing the data several times. Moreover, if the data are distributed across clients, employing standard GFlowNets leads to intensive client-server communication. To alleviate both these issues, we propose _embarrassingly parallel_ GFlowNet (EP-GFlowNet). EP-GFlowNet is a provably correct divide-and-conquer method to sample from product distributions of the form $R(\cdot) \propto R_1(\cdot) ... R_N(\cdot)$ --- e.g., in parallel or federated Bayes, where each $R_n$ is a local posterior defined on a data partition. First, in parallel, we train a local GFlowNet targeting each $R_n$ and send the resulting models to the server. Then, the server learns a global GFlowNet by enforcing our newly proposed _aggregating balance_ condition, requiring a single communication step. Importantly, EP-GFlowNets can also be applied to multi-objective optimization and model reuse. Our experiments illustrate the effectiveness of EP-GFlowNets on multiple tasks, including parallel Bayesian phylogenetics, multi-objective multiset and sequence generation, and federated Bayesian structure learning.

Current literature, aiming to surpass the "Chain-of-Thought" approach, often resorts to external modi operandi involving halting, modifying, and then resuming the generation process to boost Large Language Models' (LLMs) reasoning capacities. Due to their *myopic perspective*, they escalate the number of query requests, leading to increased costs, memory, and computational overheads. Addressing this, we propose the *Algorithm of Thoughts*---a novel strategy that propels LLMs through algorithmic reasoning pathways. By employing algorithmic examples fully in-context, this overarching view of the whole process exploits the innate recurrence dynamics of LLMs, expanding their idea exploration with merely one or a few queries. Our technique outperforms earlier single-query methods and even more recent multi-query strategies that employ an extensive tree search algorithms while using significantly fewer tokens. Intriguingly, our results suggest that instructing an LLM using an algorithm can lead to performance surpassing that of the algorithm itself, hinting at LLM's inherent ability to weave its intuition into optimized searches. We probe into the underpinnings of our method's efficacy and its nuances in application. The code and related content can be found in: https://algorithm-of-thoughts.github.io

Knowledge distillation is a widely adopted model compression technique that has been successfully applied to object detection. In feature distillation, it is common practice for the student model to imitate the feature responses of the teacher model, with the underlying objective of improving its own abilities by reducing the disparity with the teacher. However, it is crucial to recognize that the disparities between the student and teacher are inconsistent, highlighting their varying abilities. In this paper, we explore the inconsistency in the disparity between teacher and student feature maps and analyze their impact on the efficiency of the distillation. We find that regions with varying degrees of difference should be treated separately, with different distillation constraints applied accordingly. We introduce our distillation method called Disparity Feature Distillation(DFD). The core idea behind DFD is to apply different treatments to regions with varying learning difficulties, simultaneously incorporating leniency and strictness. It enables the student to better assimilate the teacher’s knowledge. Through extensive experiments, we demonstrate the effectiveness of our proposed DFD in achieving significant improvements. For instance, when applied to detectors based on ResNet50 such as RetinaNet, FasterRCNN, and RepPoints, our method enhances their mAP from 37.4%, 38.4%, 38.6% to 41.7%, 42.4%, 42.7%, respectively. Our approach also demonstrates substantial improvements on YOLO and ViT-based models. The code is available at https://github.com/luckin99/DFD.

In many image classification applications, the number of labeled training images is limited, which leads to model overfitting. To mitigate the lack of training data, deep generative models have been leveraged to generate synthetic training data. However, existing methods generate data for individual classes based on how much training data they have without considering their actual data needs. To address this limitation, we propose needs-aware image generation, which automatically identifies the different data needs of individual classes based on their classification performance and divides a limited data generation budget into these classes according to their needs. We propose a multi-level optimization based framework which performs four learning stages in an end-to-end manner. Experiments on both imbalanced and balanced classification datasets demonstrate the effectiveness of our proposed method.