ICLR 2023 Papers

Despite their remarkable achievement, gigantic transformers encounter significant drawbacks, including exorbitant computational and memory footprints during training, as well as severe collapse evidenced by a high degree of parameter redundancy. Sparsely-activated Mixture-of-Experts (SMoEs) have shown promise to mitigate the issue of training efficiency, yet they are prone to (1) $\textit{redundant experts}$ due to representational collapse; and (2) $\textit{poor expert scalability for inference and downstream fine-tuning}$, primarily due to overfitting of the learned routing policy to the number of activated experts during training. As recent research efforts are predominantly focused on improving routing policies to encourage expert specializations, this work focuses on $\textit{exploring the overlooked scalability bottleneck of SMoEs}$ and leveraging it to effectively $\textbf{scale dense transformers}$. To this end, we propose a new plug-and-play training framework, $\textbf{SMoE-Dropout}$, to enable scaling transformers to better accuracy in their full capacity without collapse. Specifically, SMoE-Dropout consists of a $\textit{randomly initialized and fixed}$ router network to activate experts and gradually increases the activated expert number as training progresses over time. Transformers trained by SMoE-Dropout naturally exhibit a $\textbf{``self-slimmable”}$ property subject to resource availability, offering smooth and consistent performance boosts with an increase in activated experts during inference or fine-tuning. Our extensive experiments across diverse transformer architectures on a variety of tasks demonstrate the superior performance and substantial computation savings of SMoE-Dropout, compared to dense training baselines with equivalent parameter counts. In particular, our trained BERT outperforms its densely trained counterpart with consistent improvements of {$1.03\%$, $0.78\%$, $1.09\%$} on challenging reasoning tasks {$\texttt{ASDiv-A}$, $\texttt{MAWPS}$, $\texttt{SVAMP}$}, respectively. Codes and models are available in https://github.com/VITA-Group/Random-MoE-as-Dropout.

A central theme in federated learning (FL) is the fact that client data distributions are often not independent and identically distributed (IID), which has strong implications on the training process. While most existing FL algorithms focus on the conventional non-IID setting of class imbalance or missing classes across clients, in practice, the distribution differences could be more complex, e.g., changes in class conditional (domain) distributions. In this paper, we consider this complex case in FL wherein each client has access to only one domain distribution. For tasks such as domain generalization, most existing learning algorithms require access to data from multiple clients (i.e., from multiple domains) during training, which is prohibitive in FL. To address this challenge, we propose a federated domain translation method that generates pseudodata for each client which could be useful for multiple downstream learning tasks. We empirically demonstrate that our translation model is more resource-efficient (in terms of both communication and computation) and easier to train in an FL setting than standard domain translation methods. Furthermore, we demonstrate that the learned translation model enables use of state-of-the-art domain generalization methods in a federated setting, which enhances accuracy and robustness to increases in the synchronization period compared to existing methodology.

Hyperparameter optimization (HPO) and neural architecture search (NAS) are methods of choice to obtain the best-in-class machine learning models, but in practice they can be costly to run. When models are trained on large datasets, tuning them with HPO or NAS rapidly becomes prohibitively expensive for practitioners, even when efficient multi-fidelity methods are employed. We propose an approach to tackle the challenge of tuning machine learning models trained on large datasets with limited computational resources. Our approach, named PASHA, extends ASHA and is able to dynamically allocate maximum resources for the tuning procedure depending on the need. The experimental comparison shows that PASHA identifies well-performing hyperparameter configurations and architectures while consuming significantly fewer computational resources than ASHA.

Heteroscedastic classifiers, which learn a multivariate Gaussian distribution over prediction logits, have been shown to perform well on image classification problems with hundreds to thousands of classes. However, compared to standard classifiers, they introduce extra parameters that scale linearly with the number of classes. This makes them infeasible to apply to larger-scale problems. In addition heteroscedastic classifiers introduce a critical temperature hyperparameter which must be tuned. We propose HET-XL, a heteroscedastic classifier whose parameter count when compared to a standard classifier scales independently of the number of classes. In our large-scale settings, we show that we can remove the need to tune the temperature hyperparameter, by directly learning it on the training data. On large image classification datasets with up to 4B images and 30k classes our method requires 14X fewer additional parameters, does not require tuning the temperature on a held-out set and performs consistently better than the baseline heteroscedastic classifier. HET-XL improves ImageNet 0-shot classification in a multimodal contrastive learning setup which can be viewed as a 3.5 billion class classification problem.

This paper presents a new pre-trained language model, NewModel, which improves the original DeBERTa model by replacing mask language modeling (MLM) with replaced token detection (RTD), a more sample-efficient pre-training task. Our analysis shows that vanilla embedding sharing in ELECTRA hurts training efficiency and model performance. This is because the training losses of the discriminator and the generator pull token embeddings in different directions, creating the “tug-of-war” dynamics. We thus propose a new gradient-disentangled embedding sharing method that avoids the tug-of-war dynamics, improving both training efficiency and the quality of the pre-trained model. We have pre-trained NewModel using the same settings as DeBERTa to demonstrate its exceptional performance on a wide range of downstream natural language understanding (NLU) tasks. Taking the GLUE benchmark with eight tasks as an example, the NewModel Large model achieves a 91.37% average score, which is 1.37% over DeBERTa and 1.91% over ELECTRA, setting a new state-of-the-art (SOTA) among the models with a similar structure. Furthermore, we have pre-trained a multi-lingual model mNew-Model and observed a larger improvement over strong baselines compared to English models. For example, the mNewModel Base achieves a 79.8% zero-shot cross-lingual accuracy on XNLI and a 3.6% improvement over XLM-R Base, creating a new SOTA on this benchmark. We will make our model and code publicly available.

Mixture-of-Experts (MoE) models are promising architectures for massively multilingual neural machine translation and large language models due to the advantage of sublinear scaling. However, the training of large MoE models is usually bottlenecked by the all-to-all communication (Lepikhin et al., 2020). To reduce the communication cost, we propose SCoMoE, an MoE architecture with structured all-to-all communication, inspired by the hierarchical architecture of the communication topology. SCoMoE encourages data to be communicated across devices through fast intra-accelerator/node communication channels, reducing communication throughput in the slow inter-node communication channel. We slice the data on the sequence dimension (SCoMoE-Seq) into three communication groups and project the data on the feature dimension (SCoMoE-Feat) into low-dimensional representations. To compensate the potential performance drop caused by the routing locality in SCoMoE, we further propose a token clustering approach to aggregating related tokens from different devices before the MoE layers. The sigmoid gating in the balanced router used in the token clustering is substituted with the softmax gating with differential sorting. Experiments on bilingual and massively multilingual machine translation demonstrate that SCoMoE achieves a speedup of 1.44x over GShard with comparable performance, and substantially outperforms Gshard (2.8 BLEU) on OPUS-100 with a speedup of 1.25x.

In many retrieval systems the original high dimensional data (e.g., images) is mapped to a lower dimensional feature through a learned embedding model. The task of retrieving the most similar data from a gallery set to a given query data is performed through similarity comparison on features. When the embedding model is updated, it might produce features that are not comparable/compatible with features already in the gallery computed with the old model. Subsequently, all features in the gallery need to be re-computed using the new embedding model -- a computationally expensive process called backfilling. Recently, compatible representation learning methods have been proposed to avoid back-filling. Despite their relative success, there is an inherent trade-off between new model performance and its compatibility with the old model. In this work, we introduce FastFill: a compatible model update process using feature alignment and policy based partial backfilling to promptly elevate retrieval performance. We show that previous backfilling strategies suffer from decreased performance and demonstrate the importance of both the training objective and the ordering in online partial backfilling. We propose a new training method for feature alignment between old and new embedding models using uncertainty estimation. Compared to previous works, we obtain significantly improved backfilling results on a variety of datasets: mAP on ImageNet (+4.4%), Places-365 (+2.7%), and VGG-Face2 (+1.3%). Further, we demonstrate that when updating a biased model with FastFill, the minority subgroup accuracy gap promptly vanishes with a small fraction of partial backfilling.

Many empirical studies have demonstrated the performance benefits of conditional computation in neural networks, including reduced inference time and power consumption. We study the fundamental limits of neural conditional computation from the perspective of memorization capacity. For Rectified Linear Unit (ReLU) networks without conditional computation, it is known that memorizing a collection of $n$ input-output relationships can be accomplished via a neural network with $O(\sqrt{n})$ neurons. Calculating the output of this neural network can be accomplished using $O(\sqrt{n})$ elementary arithmetic operations of additions, multiplications and comparisons for each input. Using a conditional ReLU network, we show that the same task can be accomplished using only $O(\log n)$ operations per input. This represents an almost exponential improvement as compared to networks without conditional computation. We also show that the $\Theta(\log n)$ rate is the best possible. Our achievability result utilizes a general methodology to synthesize a conditional network out of an unconditional network in a computationally-efficient manner, bridging the gap between unconditional and conditional architectures.

The unprecedented rate at which the sizes of machine learning (ML) models are growing necessitates novel approaches to enable efficient and scalable solutions. We contribute to this line of work by studying a novel version of the Budgeted Correlation Clustering problem (\bcc) where along with a limited number of queries to an expensive oracle for node similarities (e.g. a large ML model), we have unlimited access to a cheaper but less accurate second oracle. Our formulation is inspired by many practical scenarios where coarse approximations of the expensive similarity metric can be efficiently obtained via weaker models. We develop a theoretically motivated algorithm in this setting that leverages the cheap oracle to judiciously query the strong oracle while maintaining high clustering quality. We empirically demonstrate gains in query minimization and clustering metrics on a variety of datasets with diverse strong and cheap oracles. Most notably, we demonstrate a practical application in text clustering based on expensive cross-attention language models by showing that cheaper (but weaker) embedding-based models can be leveraged to substantially reduce the number of inference calls to the former.

Differentially private deep learning has recently witnessed advances in computational efficiency and privacy-utility trade-off. We explore whether further improvements along the two axes are possible and provide affirmative answers leveraging two instantiations of \emph{group-wise clipping}. To reduce the compute time overhead of private learning, we show that \emph{per-layer clipping}, where the gradient of each neural network layer is clipped separately, allows clipping to be performed in conjunction with backpropagation in differentially private optimization. This results in private learning that is as memory-efficient and almost as fast per training update as non-private learning for many workflows of interest. While per-layer clipping with constant thresholds tends to underperform standard flat clipping, per-layer clipping with adaptive thresholds matches or outperforms flat clipping under given training epoch constraints, hence attaining similar or better task performance within less wall time. To explore the limits of scaling (pretrained) models in differentially private deep learning, we privately fine-tune the 175 billion-parameter GPT-3. We bypass scaling challenges associated with clipping gradients that are distributed across multiple devices with \emph{per-device clipping} that clips the gradient of each model piece separately on its host device. Privately fine-tuning GPT-3 with per-device clipping achieves a task performance at $\epsilon=1$ better than what is attainable by non-privately fine-tuning the largest GPT-2 on a summarization task.

How can we design protein sequences folding into the desired structures effectively and efficiently? AI methods for structure-based protein design have attracted increasing attention in recent years; however, few methods can simultaneously improve the accuracy and efficiency due to the lack of expressive features and autoregressive sequence decoder. To address these issues, we propose PiFold, which contains a novel residue featurizer and PiGNN layers to generate protein sequences in a one-shot way with improved recovery. Experiments show that PiFold could achieve 51.66\% recovery on CATH 4.2, while the inference speed is 70 times faster than the autoregressive competitors. In addition, PiFold achieves 58.72\% and 60.42\% recovery scores on TS50 and TS500, respectively. We conduct comprehensive ablation studies to reveal the role of different types of protein features and model designs, inspiring further simplification and improvement. The PyTorch code is available at \href{https://github.com/A4Bio/PiFold}{GitHub}.

Deep neural networks (DNNs) are vulnerable to backdoor attacks, where adversaries embed a hidden backdoor trigger during the training process for malicious prediction manipulation. These attacks pose great threats to the applications of DNNs under the real-world machine learning as a service (MLaaS) setting, where the deployed model is fully black-box while the users can only query and obtain its predictions. Currently, there are many existing defenses to reduce backdoor threats. However, almost all of them cannot be adopted in MLaaS scenarios since they require getting access to or even modifying the suspicious models. In this paper, we propose a simple yet effective black-box input-level backdoor detection, called SCALE-UP, which requires only the predicted labels to alleviate this problem. Specifically, we identify and filter malicious testing samples by analyzing their prediction consistency during the pixel-wise amplification process. Our defense is motivated by an intriguing observation (dubbed \emph{scaled prediction consistency}) that the predictions of poisoned samples are significantly more consistent compared to those of benign ones when amplifying all pixel values. Besides, we also provide theoretical foundations to explain this phenomenon. Extensive experiments are conducted on benchmark datasets, verifying the effectiveness and efficiency of our defense and its resistance to potential adaptive attacks. Our codes are available at \url{https://github.com/JunfengGo/SCALE-UP}.

Since the introduction of Transformers, researchers have tackled the notoriously expensive quadratic complexity problem. While significant computational efficiency improvements have been achieved, they come at the cost of reduced accuracy trade-offs. In this paper, we propose Composite Slice Transformer (CST), a Transformer-based network equipped with a composition of multi-scale multi-range attentions, boosting both efficiency and modeling capability. After stacking fixed-length slices of the input sequence, each layer in CST performs a pair of fine-and-coarse-grained attentions with short-long ranges in a sequential manner, coupled with volatile instant positional embedding, enabling efficient token interactions {\em and} improving expressiveness of the model. In addition to significantly reduced $O(NL+N^2/L^2)$ complexity for sequence length $N$ and slice length $L$, CST achieves superior performance on a variety of tasks. We show that CST surpasses recently published efficient Transformers on the Long Range Arena benchmark, demonstrating the bidirectional long-range dependency modeling capability of our model. It outperforms the standard Transformer by a margin of $6.9$\% in average accuracy across the five classification tasks of the benchmark, while being of complexity comparable to other efficient transformers. Furthermore, on the word-level autoregressive language modeling task with the WikiText-103 dataset, CST performs competitively against the Transformer model with only $2$\% gap in the test perplexity while outperforming other efficient Transformers.

Distilling from the feature maps can be fairly effective for dense prediction tasks since both the feature discriminability and localization information can be well transferred. However, not every pixel contributes equally to the performance, and a good student should learn from what really matters to the teacher. In this paper, we introduce a learnable embedding dubbed receptive token to locate the pixels of interests (PoIs) in the feature map, with a distillation mask generated via pixel-wise attention. Then the masked distillation will be performed via the pixel-wise reconstruction. In this way, a distillation mask refers to a pattern of pixel dependencies. We thus adopt multiple receptive tokens to investigate more sophisticated and informative pixel dependencies within feature maps to enhance the distillation. To obtain a group of masks, the receptive tokens are learned via the regular task loss but with teacher fixed, and we also leverage a Dice loss to enrich the diversity of obtained masks. Our method dubbed MasKD is simple and practical, and needs no priors of ground-truth labels, which can apply to various dense prediction tasks. Experiments show that our MasKD can achieve state-of-the-art performance consistently on object detection and semantic segmentation benchmarks.

Fine-tuning large pre-trained language models on downstream tasks has become an important paradigm in NLP. However, common practice fine-tunes all of the parameters in a pre-trained model, which becomes prohibitive when a large number of downstream tasks are present. Therefore, many fine-tuning methods are proposed to learn incremental updates of pre-trained weights in a parameter efficient way, e.g., low-rank increments. These methods often evenly distribute the budget of incremental updates across all pre-trained weight matrices, and overlook the varying importance of different weight parameters. As a consequence, the fine-tuning performance is suboptimal. To bridge this gap, we propose AdaLoRA, which adaptively allocates the parameter budget among weight matrices according to their importance score. In particular, AdaLoRA parameterizes the incremental updates in the form of singular value decomposition. Such a novel approach allows us to effectively prune the singular values of unimportant updates, which is essentially to reduce their parameter budget but circumvent intensive exact SVD computations. We conduct extensive experiments with several pre-trained models on natural language processing, question answering, and natural language generation to validate the effectiveness of AdaLoRA. Results demonstrate that AdaLoRA manifests notable improvement over baselines, especially in the low budget settings. Our code is publicly available at https://github.com/QingruZhang/AdaLoRA .

3D convolution neural networks (CNNs) have been the prevailing option for video recognition. To capture the temporal information, 3D convolutions are computed along the sequences, leading to cubically growing and expensive computations. To reduce the computational cost, previous methods resort to manually designed 3D/2D CNN structures with approximations or automatic search, which sacrifice the modeling ability or make training time-consuming. In this work, we propose to automatically design efficient 3D CNN architectures via a novel training-free neural architecture search approach tailored for 3D CNNs considering the model complexity. To measure the expressiveness of 3D CNNs efficiently, we formulate a 3D CNN as an information system and derive an analytic entropy score, based on the Maximum Entropy Principle. Specifically, we propose a spatio-temporal entropy score (STEntr-Score) with a refinement factor to handle the discrepancy of visual information in spatial and temporal dimensions, through dynamically leveraging the correlation between the feature map size and kernel size depth-wisely. Highly efficient and expressive 3D CNN architectures, i.e., entropy-based 3D CNNs (E3D family), can then be efficiently searched by maximizing the STEntr-Score under a given computational budget, via an evolutionary algorithm without training the network parameters. Extensive experiments on Something-Something V1&V2 and Kinetics400 demonstrate that the E3D family achieves state-of-the-art performance with higher computational efficiency.

Recurrent neural networks (RNNs) are well suited for solving sequence tasks in resource-constrained systems due to their expressivity and low computational requirements. However, there is still a need to bridge the gap between what RNNs are capable of in terms of efficiency and performance and real-world application requirements. The memory and computational requirements arising from propagating the activations of all the neurons at every time step to every connected neuron, together with the sequential dependence of activations, contribute to the inefficiency of training and using RNNs. We propose a solution inspired by biological neuron dynamics that makes the communication between RNN units sparse and discrete. This makes the backward pass with backpropagation through time (BPTT) computationally sparse and efficient as well. We base our model on the gated recurrent unit (GRU), extending it with units that emit discrete events for communication triggered by a threshold so that no information is communicated to other units in the absence of events. We show theoretically that the communication between units, and hence the computation required for both the forward and backward passes, scales with the number of events in the network. Our model achieves efficiency without compromising task performance, demonstrating competitive performance compared to state-of-the-art recurrent network models in real-world tasks, including language modeling. The dynamic activity sparsity mechanism also makes our model well suited for novel energy-efficient neuromorphic hardware. Code is available at https://github.com/KhaleelKhan/EvNN/.

Knowledge distillation is one of the primary methods of transferring knowledge from large to small models. However, it requires massive task-specific data, which may not be plausible in many real-world applications. Data augmentation methods such as representation interpolation, token replacement, or augmentation with models are applied to tackle this problem. However, these data augmentation methods either potentially cause shifts in decision boundaries (representation interpolation), are not expressive enough (token replacement), or introduce too much computational overhead (augmentation with models). To this end, we propose AugPro (Augmentation with Projection), an effective and efficient data augmentation method for distillation. Our method builds on top of representation interpolation augmentation methods to maintain the diversity of expressions and converts the augmented data to tokens to avoid shifting decision boundaries. It uses simple operations that come with little computational overhead. The results on multiple GLUE tasks show that our methods can improve distillation performance by a large margin at a low time cost.

One main challenge in federated learning is the large communication cost of exchanging weight updates from clients to the server at each round. While prior work has made great progress in compressing the weight updates through gradient compression methods, we propose a radically different approach that does not update the weights at all. Instead, our method freezes the weights at their initial \emph{random} values and learns how to sparsify the random network for the best performance. To this end, the clients collaborate in training a \emph{stochastic} binary mask to find the optimal sparse random network within the original one. At the end of the training, the final model is a sparse network with random weights -- or a subnetwork inside the dense random network. We show improvements in accuracy, communication (less than $1$ bit per parameter (bpp)), convergence speed, and final model size (less than $1$ bpp) over relevant baselines on MNIST, EMNIST, CIFAR-10, and CIFAR-100 datasets, in the low bitrate regime.