ICML 2024 Papers

In this paper, we introduce a novel model compression technique named Joint Sparsification and Quantization (JSQ), explicitly tailored for large language models (LLMs). Traditional methods employ either sparsification or quantization individually to compress LLMs, leading to performance degradation at high compression ratios. In contrast, our JSQ approach integrates sparsification and quantization cohesively. As sparsification tend to preserve outliers that is harmful to quantization, we introduce a novel sparsity metric to serves as a bridge between the sparsification and quantization. Moreover, it is proven outliers in LLMs have significant impact but harmful to compression. Current solutions are highly coupled with quantization process, which is not helpful to sparsification. To this end, we also introduce a search-based activation editor to automatically eliminate relatively useless outliers. Comprehensive experiments across various datasets and architectures affirm the efficacy of our JSQ framework. Notably, our JSQ achieves 7.96$\times$ computation reduction without crashing for the representative model LLaMA. This accomplishment stands in stark contrast to the limitations of most state-of-the-art LLM compression methods, which typically fail under such extreme compression ratios. Our code is released at https://github.com/uanu2002/JSQ.

Deep neural policies have recently been installed in a diverse range of settings, from biotechnology to automated financial systems. However, the utilization of deep neural networks to approximate the value function leads to concerns on the decision boundary stability, in particular, with regard to the sensitivity of policy decision making to indiscernible, non-robust features due to highly non-convex and complex deep neural manifolds. These concerns constitute an obstruction to understanding the reasoning made by deep neural policies, and their foundational limitations. Hence, it is crucial to develop techniques that aim to understand the sensitivities in the learnt representations of neural network policies. To achieve this we introduce a theoretically founded method that provides a systematic analysis of the unstable directions in the deep neural policy decision boundary across both time and space. Through experiments in the Arcade Learning Environment (ALE), we demonstrate the effectiveness of our technique for identifying correlated directions of instability, and for measuring how sample shifts remold the set of sensitive directions in the neural policy landscape. Most importantly, we demonstrate that state-of-the-art robust training techniques yield learning of disjoint unstable directions, with dramatically larger oscillations over time, when compared to standard training. We believe our results reveal the fundamental properties of the decision process made by reinforcement learning policies, and can help in constructing reliable and robust deep neural policies.

Multi-timestep simulation of brain-inspired Spiking Neural Networks (SNNs) boost memory requirements during training and increase inference energy cost. Current training methods cannot simultaneously solve both training and inference dilemmas. This work proposes a novel Temporal Reversible architecture for SNNs (T-RevSNN) to jointly address the training and inference challenges by altering the forward propagation of SNNs. We turn off the temporal dynamics of most spiking neurons and design multi-level temporal reversible interactions at temporal turn-on spiking neurons, resulting in a $\mathcal{O}(L)$ training memory. Combined with the temporal reversible nature, we redesign the input encoding and network organization of SNNs to achieve $\mathcal{O}(1)$ inference energy cost. Then, we finely adjust the internal units and residual connections of the basic SNN block to ensure the effectiveness of sparse temporal information interaction. T-RevSNN achieves excellent accuracy on ImageNet, while the memory efficiency, training time acceleration and inference energy efficiency can be significantly improved by $8.6 \times$, $2.0 \times$ and $1.6 \times$, respectively. This work is expected to break the technical bottleneck of significantly increasing memory cost and training time for large-scale SNNs while maintaining both high performance and low inference energy cost.

How do neural networks extract patterns from pixels? Feature visualizations attempt to answer this important question by visualizing highly activating patterns through optimization. Today, visualization methods form the foundation of our knowledge about the internal workings of neural networks, as a type of mechanistic interpretability. Here we ask: How reliable are feature visualizations? We start our investigation by developing network circuits that trick feature visualizations into showing arbitrary patterns that are completely disconnected from normal network behavior on natural input. We then provide evidence for a similar phenomenon occurring in standard, unmanipulated networks: feature visualizations are processed very differently from standard input, casting doubt on their ability to "explain" how neural networks process natural images. This can be used as a sanity check for feature visualizations. We underpin our empirical findings by theory proving that the set of functions that can be reliably understood by feature visualization is extremely small and does not include general black-box neural networks. Therefore, a promising way forward could be the development of networks that enforce certain structures in order to ensure more reliable feature visualizations.

Differentiable annealed importance sampling (DAIS), proposed by Geffner & Domke (2021) and Zhang et al. (2021), allows optimizing, among others, over the initial distribution of AIS. In this paper, we show that, in the limit of many transitions, DAIS minimizes the symmetrized KL divergence (Jensen-Shannon divergence) between the initial and target distribution. Thus, DAIS can be seen as a form of variational inference (VI) in that its initial distribution is a parametric fit to an intractable target distribution. We empirically evaluate the usefulness of the initial distribution as a variational distribution on synthetic and real-world data, observing that it often provides more accurate uncertainty estimates than standard VI (optimizing the reverse KL divergence), importance weighted VI, and Markovian score climbing (optimizing the forward KL divergence).

Regression trees are a human-comprehensible machine-learning model that can represent complex relationships. They are typically trained using greedy heuristics because computing optimal regression trees is NP-hard. Contrary to this standard practice, we consider optimal methods and improve the scalability of optimal methods by developing three new dynamic programming approaches. First, we improve the performance of a piecewise constant regression tree method using a special algorithm for trees of depth two. Second, we provide the first optimal dynamic programming method for piecewise multiple linear regression. Third, we develop the first optimal method for piecewise simple linear regression, for which we also provide a special algorithm for trees of depth two. The experimental results show that our methods improve scalability by one or more orders of magnitude over the state-of-the-art optimal methods while performing similarly or better in out-of-sample performance.

The current state-of-the-art theoretical analysis of Actor-Critic (AC) algorithms significantly lags in addressing the practical aspects of AC implementations. This crucial gap needs bridging to bring the analysis in line with practical implementations of AC. To address this, we advocate for considering the MMCLG criteria: **M**ulti-layer neural network parametrization for actor/critic, **M**arkovian sampling, **C**ontinuous state-action spaces, the performance of the **L**ast iterate, and **G**lobal optimality. These aspects are practically significant and have been largely overlooked in existing theoretical analyses of AC algorithms. In this work, we address these gaps by providing the first comprehensive theoretical analysis of AC algorithms that encompasses all five crucial practical aspects (covers MMCLG criteria). We establish global convergence sample complexity bounds of $\tilde{\mathcal{O}}\left( \epsilon^{-3} \right)$. We achieve this result through our novel use of the weak gradient domination property of MDP's and our unique analysis of the error in critic estimation.

Exploration remains a critical issue in deep reinforcement learning for an agent to attain high returns in unknown environments. Although the prevailing exploration Random Network Distillation (RND) algorithm has been demonstrated to be effective in numerous environments, it often needs more discriminative power in bonus allocation. This paper highlights the ``bonus inconsistency'' issue within RND, pinpointing its primary limitation. To address this issue, we introduce the Distributional RND (DRND), a derivative of the RND. DRND enhances the exploration process by distilling a distribution of random networks and implicitly incorporating pseudo counts to improve the precision of bonus allocation. This refinement encourages agents to engage in more extensive exploration. Our method effectively mitigates the inconsistency issue without introducing significant computational overhead. Both theoretical analysis and experimental results demonstrate the superiority of our approach over the original RND algorithm. Our method excels in challenging online exploration scenarios and effectively serves as an anti-exploration mechanism in D4RL offline tasks. Our code is publicly available at https://github.com/yk7333/DRND.

Suppose a self-driving car is crashing into pedestrians, or a chatbot is instructing its users to conduct criminal wrongdoing; the stakeholders of such products will undoubtedly want to patch these catastrophic errors as soon as possible. To address such concerns, *Model Editing:* the study of efficiently patching model behaviors without significantly altering their general performance, has seen considerable activity, with hundreds of editing techniques developed in various domains such as CV and NLP. However, **the graph learning community has objectively fallen behind with only a few Graph Neural Network-compatible — and just one GNN-specific — model editing methods available**, where all of which are limited in their practical scope. We argue that the impracticality of these methods lies in their lack of *Sequential Editing Robustness:* the ability to edit multiple errors sequentially, and therefore fall short in effectiveness, as this approach mirrors how errors are discovered and addressed in the real world. In this paper, we delve into the specific reasons behind the difficulty of editing GNNs in succession and observe the root cause to be model overfitting. We subsequently propose a simple yet effective solution — SEED-GNN — by leveraging overfit-prevention techniques in a GNN-specific context to derive the first and only GNN model editing method that scales practically. Additionally, we formally frame the task paradigm of GNN editing and hope to inspire future research in this crucial but currently overlooked field. Please refer to our [GitHub repository](https://github.com/henryzhongsc/gnn_editing) for code and checkpoints.

Robotic manipulation tasks often span over long horizons and encapsulate multiple subtasks with different skills. Learning policies directly from long-horizon demonstrations is challenging without intermediate keyframes guidance and corresponding skill annotations. Existing approaches for keyframe identification often struggle to offer reliable decomposition for low accuracy and fail to provide semantic relevance between keyframes and skills. For this, we propose a unified **K**eyframe **I**dentifier and **S**kill **A**notator (**KISA**) that utilizes pretrained visual-language representations for precise and interpretable decomposition of unlabeled demonstrations. Specifically, we develop a simple yet effective temporal enhancement module that enriches frame-level representations with expanded receptive fields to capture semantic dynamics at the video level. We further propose coarse contrastive learning and fine-grained monotonic encouragement to enhance the alignment between visual representations from keyframes and language representations from skills. The experimental results across three benchmarks demonstrate that KISA outperforms competitive baselines in terms of accuracy and interpretability of keyframe identification. Moreover, KISA exhibits robust generalization capabilities and the flexibility to incorporate various pretrained representations.

Although deep learning models have taken on commercial and political relevance, key aspects of their training and operation remain poorly understood. This has sparked interest in science of deep learning projects, many of which require large amounts of time, money, and electricity. But how much of this research really needs to occur at scale? In this paper, we introduce MNIST-1D: a minimalist, procedurally generated, low-memory, and low-compute alternative to classic deep learning benchmarks. Although the dimensionality of MNIST-1D is only 40 and its default training set size only 4000, MNIST-1D can be used to study inductive biases of different deep architectures, find lottery tickets, observe deep double descent, metalearn an activation function, and demonstrate guillotine regularization in self-supervised learning. All these experiments can be conducted on a GPU or often even on a CPU within minutes, allowing for fast prototyping, educational use cases, and cutting-edge research on a low budget.

Conventional Federated Learning (FL) involves collaborative training of a global model while maintaining user data privacy. One of its branches, decentralized FL, is a serverless network that allows clients to own and optimize different local models separately, which results in saving management and communication resources. Despite the promising advancements in decentralized FL, it may reduce model generalizability due to lacking a global model. In this scenario, managing data and model heterogeneity among clients becomes a crucial problem, which poses a unique challenge that must be overcome: *How can every client's local model learn generalizable representation in a decentralized manner?* To address this challenge, we propose a novel **De**centralized FL technique by introducing **S**ynthetic **A**nchors, dubbed as DeSA. Based on the theory of domain adaptation and Knowledge Distillation (KD), we theoretically and empirically show that synthesizing global anchors based on raw data distribution facilitates mutual knowledge transfer. We further design two effective regularization terms for local training: *1) REG loss* that regularizes the distribution of the client's latent embedding with the anchors and *2) KD loss* that enables clients to learn from others. Through extensive experiments on diverse client data distributions, we showcase the effectiveness of DeSA in enhancing both inter- and intra-domain accuracy of each client. The implementation of DeSA can be found at: https://github.com/ubc-tea/DESA

This paper describes MAIA, a Multimodal Automated Interpretability Agent. MAIA is a system that uses neural models to automate neural model understanding tasks like feature interpretation and failure mode discovery. It equips a pre-trained vision-language model with a set of tools that support iterative experimentation on subcomponents of other models to explain their behavior. These include tools commonly used by human interpretability researchers: for synthesizing and editing inputs, computing maximally activating exemplars from real-world datasets, and summarizing and describing experimental results. Interpretability experiments proposed by MAIA compose these tools to describe and explain system behavior. We evaluate applications of MAIA to computer vision models. We first characterize MAIA’s ability to describe (neuron-level) features in learned representations of images. Across several trained models and a novel dataset of synthetic vision neurons with paired ground-truth descriptions, MAIA produces descriptions comparable to those generated by expert human experimenters. We then show that MAIA can aid in two additional interpretability tasks: reducing sensitivity to spurious features, and automatically identifying inputs likely to be mis-classified.

Deep neural networks come in many sizes and architectures. The choice of architecture, in conjunction with the dataset and learning algorithm, is commonly understood to affect the learned neural representations. Yet, recent results have shown that different architectures learn representations with striking qualitative similarities. Here we derive an effective theory of representation learning under the assumption that the encoding map from input to hidden representation and the decoding map from representation to output are arbitrary smooth functions. This theory schematizes representation learning dynamics in the regime of complex, large architectures, where hidden representations are not strongly constrained by the parametrization. We show through experiments that the effective theory describes aspects of representation learning dynamics across a range of deep networks with different activation functions and architectures, and exhibits phenomena similar to the “rich” and “lazy” regime. While many network behaviors depend quantitatively on architecture, our findings point to certain behaviors that are widely conserved once models are sufficiently flexible.

We focus on constrained, $L$-smooth, potentially stochastic and nonconvex-nonconcave min-max problems either satisfying $\rho$-cohypomonotonicity or admitting a solution to the $\rho$-weakly Minty Variational Inequality (MVI), where larger values of the parameter $\rho>0$ correspond to a greater degree of nonconvexity. These problem classes include examples in two player reinforcement learning, interaction dominant min-max problems, and certain synthetic test problems on which classical min-max algorithms fail. It has been conjectured that first-order methods can tolerate a value of $\rho$ no larger than $\frac{1}{L}$, but existing results in the literature have stagnated at the tighter requirement $\rho < \frac{1}{2L}$. With a simple argument, we obtain optimal or best-known complexity guarantees with cohypomonotonicity or weak MVI conditions for $\rho < \frac{1}{L}$. First main insight for the improvements in the convergence analyses is to harness the recently proposed *conic nonexpansiveness* property of operators. Second, we provide a refined analysis for inexact Halpern iteration that relaxes the required inexactness level to improve some state-of-the-art complexity results even for constrained stochastic convex-concave min-max problems. Third, we analyze a stochastic inexact Krasnosel'skii-Mann iteration with a multilevel Monte Carlo estimator when the assumptions only hold with respect to a solution.

Aligning large language models (LLMs) with human values is imperative to mitigate potential adverse effects resulting from their misuse. Drawing from the sociological insight that acknowledging all parties' concerns is a key factor in shaping human values, this paper proposes a novel direction to align LLMs by themselves: social scene simulation. To achieve this, we present MATRIX, a novel social scene simulator that emulates realistic scenes around a user's input query, enabling the LLM to take social consequences into account before responding. MATRIX serves as a virtual rehearsal space, akin to a Monopolylogue, where the LLM performs diverse roles related to the query and practice by itself. To inject this alignment, we fine-tune the LLM with MATRIX-simulated data, ensuring adherence to human values without compromising inference speed. We theoretically show that the LLM with MATRIX outperforms existing methods under mild assumptions. Finally, extensive experiments validate that our method outperforms over 10 baselines across 4 benchmarks. As evidenced by 875 user ratings, our tuned 13B-size LLM exceeds GPT-4 in aligning with human values. See our project page at https://shuotang123.github.io/MATRIX.

Mixup is a highly successful technique to improve generalization by augmenting training data with combinations of random pairs. Selective mixup is a family of methods that apply mixup to specific pairs e.g. combining examples across classes or domains. Despite remarkable performance on benchmarks with distribution shifts, these methods are still poorly understood. We find that an overlooked aspect of selective mixup explains some of its success in a completely new light. The non-random selection of pairs affects the training distribution and improves generalization by means completely unrelated to the mixing. For example in binary classification, mixup across classes implicitly resamples the data to uniform class distribution - a classical solution to label shift. We verify empirically that this resampling explains some of the improvements reported in prior work. Theoretically, the effect relies on a ``regression toward the mean'', an accidental property we find in several datasets. Outcomes. We now better understand why selective mixup works. This lets us predict a yet-unknown failure mode and conditions where the method is detrimental. We also use the equivalence with resampling to design better variants that combine mixing and resampling effects.

Machine learning (ML) datasets, often perceived as neutral, inherently encapsulate abstract and disputed social constructs. Dataset curators frequently employ value-laden terms such as diversity, bias, and quality to characterize datasets. Despite their prevalence, these terms lack clear definitions and validation. Our research explores the implications of this issue by analyzing "diversity" across 135 image and text datasets. Drawing from social sciences, we apply principles from measurement theory to identify considerations and offer recommendations for conceptualizing, operationalizing, and evaluating diversity in datasets. Our findings have broader implications for ML research, advocating for a more nuanced and precise approach to handling value-laden properties in dataset construction.

In this paper, we first present an explanation regarding the common occurrence of spikes in the training loss when neural networks are trained with stochastic gradient descent (SGD). We provide evidence that the spikes in the training loss of SGD are "catapults", an optimization phenomenon originally observed in GD with large learning rates in Lewkowycz et al. (2020). We empirically show that these catapults occur in a low-dimensional subspace spanned by the top eigenvectors of the tangent kernel, for both GD and SGD. Second, we posit an explanation for how catapults lead to better generalization by demonstrating that catapults increase feature learning by increasing alignment with the Average Gradient Outer Product (AGOP) of the true predictor. Furthermore, we demonstrate that a smaller batch size in SGD induces a larger number of catapults, thereby improving AGOP alignment and test performance.