ICML 2024 Papers

We revisit the problem of federated learning (FL) with private data from people who do not trust the server or other silos/clients. In this context, every silo (e.g. hospital) has data from several people (e.g. patients) and needs to protect the privacy of each person's data (e.g. health records), even if the server and/or other silos try to uncover this data. Inter-Silo Record-Level Differential Privacy (ISRL-DP) prevents each silo's data from being leaked, by requiring that silo $i$'s *communications* satisfy item-level differential privacy. Prior work (Lowy & Razaviyayn, 2023a) characterized the optimal excess risk bounds for ISRL-DP algorithms with *homogeneous* (i.i.d.) silo data and convex loss functions. However, two important questions were left open: 1) Can the same excess risk bounds be achieved with *heterogeneous* (non-i.i.d.) silo data? 2) Can the optimal risk bounds be achieved with *fewer communication rounds*? In this paper, we give positive answers to both questions. We provide novel ISRL-DP FL algorithms that achieve the optimal excess risk bounds in the presence of heterogeneous silo data. Moreover, our algorithms are more *communication-efficient* than the prior state-of-the-art. For smooth loss functions, our algorithm achieves the *optimal* excess risk bound and has *communication complexity that matches the non-private lower bound*. Additionally, our algorithms are more *computationally efficient* than the previous state-of-the-art.

Large language models (LLMs) are instruction followers but the performance varies under different instructions. It is challenging to create the best instruction, especially for black-box LLMs on which backpropagation is forbidden. Instead of directly optimizing the discrete instruction, we optimize a low-dimensional soft prompt applied to an open-source LLM to generate the instruction for the black-box LLM. In each optimization step of the proposed method InstructZero, a soft prompt is converted into an instruction by the open-source LLM, which is then submitted to the black-box LLM for zero-shot evaluation, whose result is sent to Bayesian optimization to produce new soft prompts improving the zero-shot performance. We evaluate InstructZero on different combinations of open-source LLMs and APIs including Vicuna and ChatGPT. InstructZero outperforms SOTA auto-instruction methods across a variety of downstream tasks.

Contrastive Language-Image Pretraining (CLIP) has achieved remarkable success, leading to rapid advancements in multimodal studies. However, CLIP faces a notable challenge in terms of *inefficient data utilization*. It relies on a single contrastive supervision for each image-text pair during representation learning, disregarding a substantial amount of valuable information that could offer richer supervision. Additionally, the retention of non-informative tokens leads to increased computational demands and time costs, particularly in CLIP's ViT image encoder. To address these issues, we propose **M**ulti-Perspective **L**anguage-**I**mage **P**retraining (**MLIP**). In MLIP, we leverage the frequency transform's sensitivity to both high and low-frequency variations, which complements the spatial domain's sensitivity limited to low-frequency variations only. By incorporating frequency transforms and token-level alignment, we expand CILP's single supervision into multi-domain and multi-level supervision, enabling a more thorough exploration of informative image features. Additionally, we introduce a token merging method guided by comprehensive semantics from the frequency and spatial domains. This allows us to merge tokens to multi-granularity tokens with a controllable compression rate to accelerate CLIP. Extensive experiments validate the effectiveness of our design.

Dataset distillation aims to synthesize a small number of images per class (IPC) from a large dataset to approximate full dataset training with minimal performance loss. While effective in very small IPC ranges, many distillation methods become less effective, even underperforming random sample selection, as IPC increases. Our examination of state-of-the-art trajectory-matching based distillation methods across various IPC scales reveals that these methods struggle to incorporate the complex, rare features of harder samples into the synthetic dataset even with the increased IPC, resulting in a persistent coverage gap between easy and hard test samples. Motivated by such observations, we introduce SelMatch, a novel distillation method that effectively scales with IPC. SelMatch uses selection-based initialization and partial updates through trajectory matching to manage the synthetic dataset's desired difficulty level tailored to IPC scales. When tested on CIFAR-10/100 and TinyImageNet, SelMatch consistently outperforms leading selection-only and distillation-only methods across subset ratios from 5% to 30%.

Uncertainty estimation is crucial for deep learning models to detect out-of-distribution (OOD) inputs. However, the naive deep learning classifiers produce uncalibrated uncertainty for OOD data. Improving the uncertainty estimation typically requires external data for OOD-aware training or considerable costs to build an ensemble. In this work, we improve on uncertainty estimation without extra OOD data or additional inference costs using an alternative *Split-Ensemble* method. Specifically, we propose a novel *subtask-splitting* ensemble training objective where a task is split into several complementary subtasks based on feature similarity. Each subtask considers part of the data as in distribution while all the rest as OOD data. Diverse submodels can therefore be trained on each subtask with OOD-aware objectives, learning generalizable uncertainty estimation. To avoid overheads, we enable low-level feature sharing among submodels, building a tree-like Split-Ensemble architecture via iterative splitting and pruning. Empirical study shows Split-Ensemble, without additional computational cost, improves accuracy over a single model by 0.8%, 1.8%, and 25.5% on CIFAR-10, CIFAR-100, and Tiny-ImageNet, respectively. OOD detection for the same backbone and in-distribution datasets surpasses a single model baseline by 2.2%, 8.1%, and 29.6% in mean AUROC, respectively.

In this paper, we investigate the empirical counterpart of Group Distributionally Robust Optimization (GDRO), which aims to minimize the maximal empirical risk across $m$ distinct groups. We formulate empirical GDRO as a *two-level* finite-sum convex-concave minimax optimization problem and develop an algorithm called ALEG to benefit from its special structure. ALEG is a double-looped stochastic primal-dual algorithm that incorporates variance reduction techniques into a modified mirror prox routine. To exploit the two-level finite-sum structure, we propose a simple group sampling strategy to construct the stochastic gradient with a smaller Lipschitz constant and then perform variance reduction for all groups. Theoretical analysis shows that ALEG achieves $\varepsilon$-accuracy within a computation complexity of $\mathcal{O}\left(\frac{m\sqrt{\bar{n}\ln{m}}}{\varepsilon}\right)$, where $\bar n$ is the average number of samples among $m$ groups. Notably, our approach outperforms the state-of-the-art method by a factor of $\sqrt{m}$. Based on ALEG, we further develop a two-stage optimization algorithm called ALEM to deal with the empirical Minimax Excess Risk Optimization (MERO) problem. The computation complexity of ALEM nearly matches that of ALEG, surpassing the rates of existing methods.

While traditional distributionally robust optimization (DRO) aims to minimize the maximal risk over a set of distributions, Agarwal & Zhang (2022) recently proposed a variant that replaces risk with *excess risk*. Compared to DRO, the new formulation—minimax excess risk optimization (MERO) has the advantage of suppressing the effect of heterogeneous noise in different distributions. However, the choice of excess risk leads to a very challenging minimax optimization problem, and currently there exists only an inefficient algorithm for empirical MERO. In this paper, we develop efficient stochastic approximation approaches which directly target MERO. Specifically, we leverage techniques from stochastic convex optimization to estimate the minimal risk of every distribution, and solve MERO as a stochastic convex-concave optimization (SCCO) problem with biased gradients. The presence of bias makes existing theoretical guarantees of SCCO inapplicable, and fortunately, we demonstrate that the bias, caused by the estimation error of the minimal risk, is under-control. Thus, MERO can still be optimized with a nearly optimal convergence rate. Moreover, we investigate a practical scenario where the quantity of samples drawn from each distribution may differ, and propose a stochastic approach that delivers *distribution-dependent* convergence rates.

Masked Video Autoencoder (MVA) approaches have demonstrated their potential by significantly outperforming previous video representation learning methods. However, they waste an excessive amount of computations and memory in predicting uninformative tokens/frames due to random masking strategies. (e.g., over 16 nodes with 128 NVIDIA A100 GPUs). To resolve this issue, we exploit the unequal information density among the patches in videos and propose EVEREST, a surprisingly efficient MVA approach for video representation learning that finds tokens containing rich motion features and discards uninformative ones during both pre-training and fine-tuning. We further present an information-intensive frame selection strategy that allows the model to focus on informative and causal frames with minimal redundancy. Our method significantly reduces the computation and memory requirements of MVA, enabling the pre-training and fine-tuning on a single machine with 8 GPUs while achieving comparable performance to computation- and memory-heavy baselines on multiple benchmarks and the uncurated Ego4D dataset. We hope that our work contributes to reducing the barrier to further research on video understanding.

On a variety of tasks, the performance of neural networks predictably improves with training time, dataset size and model size across many orders of magnitude. This phenomenon is known as a neural scaling law. Of fundamental importance is the compute-optimal scaling law, which reports the performance as a function of units of compute when choosing model sizes optimally. We analyze a random feature model trained with gradient descent as a solvable model of network training and generalization. This reproduces many observations about neural scaling laws. First, our model makes a prediction about why the scaling of performance with training time and with model size have different power law exponents. Consequently, the theory predicts an asymmetric compute-optimal scaling rule where the number of training steps are increased faster than model parameters, consistent with recent empirical observations. Second, it has been observed that early in training, networks converge to their infinite-width dynamics at a rate $1/\text{width}$ but at late time exhibit a rate $\text{width}^{-c}$, where $c$ depends on the structure of the architecture and task. We show that our model exhibits this behavior. Lastly, our theory shows how the gap between training and test loss can gradually build up over time due to repeated reuse of data.

Analog electrical networks have long been investigated as energy-efficient computing platforms for machine learning, leveraging analog physics during inference. More recently, resistor networks have sparked particular interest due to their ability to learn using local rules (such as equilibrium propagation), enabling potentially important energy efficiency gains for training as well. Despite their potential advantage, the simulations of these resistor networks has been a significant bottleneck to assess their scalability, with current methods either being limited to linear networks or relying on realistic, yet slow circuit simulators like SPICE. Assuming ideal circuit elements, we introduce a novel approach for the simulation of nonlinear resistive networks, which we frame as a quadratic programming problem with linear inequality constraints, and which we solve using a fast, exact coordinate descent algorithm. Our simulation methodology significantly outperforms existing SPICE-based simulations, enabling the training of networks up to 327 times larger at speeds 160 times faster, resulting in a 50,000-fold improvement in the ratio of network size to epoch duration. Our approach can foster more rapid progress in the simulations of nonlinear analog electrical networks.

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.

Pretraining transformers are generally time-consuming. Fully quantized training (FQT) is a promising approach to speed up pretraining. However, most FQT methods adopt a quantize-compute-dequantize procedure, which often leads to suboptimal speedup and significant performance degradation when used in transformers due to the high memory access overheads and low-precision computations. In this work, we propose Jetfire, an efficient and accurate INT8 training method specific to transformers. Our method features an INT8 data flow to optimize memory access and a per-block quantization method to maintain the accuracy of pretrained transformers. Extensive experiments demonstrate that our INT8 FQT method achieves comparable accuracy to the FP16 training baseline and outperforms the existing INT8 training works for transformers. Moreover, for a standard transformer block, our method offers an end-to-end training speedup of 1.42x and a 1.49x memory reduction compared to the FP16 baseline.

Knowledge distillation (KD) is widely used for compressing a teacher model to a smaller student model, reducing its inference cost and memory footprint while preserving model capabilities. However, current KD methods for auto-regressive sequence models (e.g., large language models) suffer from missing a standardized objective function. Moreover, the recent use of student-generated outputs to address training-inference mismatches has significantly escalated computational costs. To tackle these issues, we introduce DistiLLM, a more effective and efficient KD framework for auto-regressive language models. DistiLLM comprises two components: (1) a novel skew Kullback-Leibler divergence loss, where we unveil and leverage its theoretical properties, and (2) an adaptive off-policy approach designed to enhance the efficiency in utilizing student-generated outputs. Extensive experiments, including instruction-following tasks, demonstrate the effectiveness of DistiLLM in building high-performing student models while achieving up to 4.3$\times$ speedup compared to recent KD methods.

Sampling-based methods, e.g., Deep Ensembles and Bayesian Neural Nets have become promising approaches to improve the quality of uncertainty estimation and robust generalization. However, they suffer from a large model size and high latency at test time, which limits the scalability needed for low-resource devices and real-time applications. To resolve these computational issues, we propose Density-Softmax, a sampling-free deterministic framework via combining a density function built on a Lipschitz-constrained feature extractor with the softmax layer. Theoretically, we show that our model is the solution of minimax uncertainty risk and is distance-aware on feature space, thus reducing the over-confidence of the standard softmax under distribution shifts. Empirically, our method enjoys competitive results with state-of-the-art techniques in terms of uncertainty and robustness, while having a lower number of model parameters and a lower latency at test time.

We introduce an Outlier-Efficient Modern Hopfield Model (termed `OutEffHop`) and use it to address the outlier inefficiency problem of training gigantic transformer-based models. Our main contribution is a novel associative memory model facilitating _outlier-efficient_ associative memory retrievals. Interestingly, this memory model manifests a model-based interpretation of an outlier-efficient attention mechanism (`Softmax_1`): it is an approximation of the memory retrieval process of `OutEffHop`. Methodologically, this allows us to introduce novel outlier-efficient Hopfield layers as powerful alternatives to traditional attention mechanisms, with superior post-quantization performance. Theoretically, the Outlier-Efficient Modern Hopfield Model retains and improves the desirable properties of standard modern Hopfield models, including fixed point convergence and exponential storage capacity. Empirically, we demonstrate the efficacy of the proposed model across large-scale transformer-based and Hopfield-based models (including BERT, OPT, ViT, and STanHop-Net), benchmarking against state-of-the-art methods like `Clipped_Softmax` and `Gated_Attention`. Notably, `OutEffHop` achieves an average reduction of 22+% in average kurtosis and 26+% in the maximum infinity norm of model outputs across four models. Code is available at [GitHub](https://github.com/MAGICS-LAB/OutEffHop); future updates are on [arXiv](https://arxiv.org/abs/2404.03828).

In this work, we propose a novel optimization model termed ``sum-of-minimum" optimization. This model seeks to minimize the sum or average of $N$ objective functions over $k$ parameters, where each objective takes the minimum value of a predefined sub-function with respect to the $k$ parameters. This universal framework encompasses numerous clustering applications in machine learning and related fields. We develop efficient algorithms for solving sum-of-minimum optimization problems, inspired by a randomized initialization algorithm for the classic $k$-means (Arthur & Vassilvitskii, 2007) and Lloyd's algorithm (Lloyd, 1982). We establish a new tight bound for the generalized initialization algorithm and prove a gradient-descent-like convergence rate for generalized Lloyd's algorithm. The efficiency of our algorithms is numerically examined on multiple tasks, including generalized principal component analysis, mixed linear regression, and small-scale neural network training. Our approach compares favorably to previous ones based on simpler-but-less-precise optimization reformulations.

Despite impressive results, deep generative models require massive datasets for training, and as dataset size increases, effective evaluation metrics like precision and recall (P&R) become computationally infeasible on commodity hardware. In this paper, we address this challenge by proposing efficient P&R (eP&R) metrics that give almost identical results as the original P&R but with much lower computational costs. Specifically, we identify two redundancies in the original P&R: i) redundancy in ratio computation and ii) redundancy in manifold inside/outside identification. We find both can be effectively removed via hubness-aware sampling, which extracts representative elements from synthetic/real image samples based on their hubness values, i.e., the number of times a sample becomes a k-nearest neighbor to others in the feature space. Thanks to the insensitivity of hubness-aware sampling to exact k-nearest neighbor (k-NN) results, we further improve the efficiency of our eP&R metrics by using approximate k-NN methods. Extensive experiments show that our eP&R matches the original P&R but is far more efficient in time and space. Our code is available at: https://github.com/Byronliang8/Hubness_Precision_Recall

Transformers with linear attention allow for efficient parallel training but can simultaneously be formulated as an RNN with 2D (matrix-valued) hidden states, thus enjoying linear-time inference complexity. However, linear attention generally underperforms ordinary softmax attention. Moreover, current implementations of linear attention lack I/O-awareness and are thus slower than highly optimized implementations of softmax attention. This work describes a hardware-efficient algorithm for linear attention that trades off memory movement against parallelizability. The resulting implementation, dubbed FlashLinearAttention, is faster than FlashAttention-2 as a standalone layer even on short sequence lengths (e.g., 1K). We then generalize this algorithm to a more expressive variant of linear attention with data-dependent gates. When used as a replacement for the standard attention layer in Transformers, the resulting gated linear attention (GLA) Transformer is found to perform competitively against the LLaMA-architecture Transformer as well recent linear-time-inference baselines such as RetNet and Mamba on moderate-scale language modeling experiments. GLA Transformer is especially effective at length generalization, enabling a model trained on 2K to generalize to sequences longer than 20K without significant perplexity degradations. For training speed, the GLA Transformer has higher throughput than a similarly-sized Mamba model.

The increasing size of large language models (LLMs) traditionally requires low-precision integer formats to meet strict latency and power demands. Yet recently, alternative formats such as Normal Float (NF4) have increased model accuracy at the cost of increased chip area. In this work, we first conduct a large-scale analysis of LLM weights and activations across 30 networks and conclude that most distributions follow a Student's t-distribution. We then derive a new theoretically optimal format, Student Float (SF4), that improves over NF4 across modern LLMs, for example increasing the average accuracy on LLaMA2-7B by 0.76% across tasks. Using this format as a high-accuracy reference, we then propose augmenting E2M1 with two variants of *supernormal* support for higher model accuracy. Finally, we explore the quality and efficiency frontier across 11 datatypes by evaluating their model accuracy and hardware complexity. We discover a Pareto curve composed of INT4, E2M1, and E2M1 with supernormal support, which offers a continuous tradeoff between model accuracy and chip area. For example, E2M1 with supernormal support increases the accuracy of Phi-2 by up to 2.19% with 1.22% area overhead, enabling more LLM-based applications to be run at four bits. The supporting code is hosted at https://github.com/cornell-zhang/llm-datatypes.