ICLR 2023 Papers

Edge devices provide inference on predictive tasks to many end-users. However, deploying deep neural networks that achieve state-of-the-art accuracy on these devices is infeasible due to edge resource constraints. Nevertheless, cloud-only processing, the de-facto standard, is also problematic, since uploading large amounts of data imposes severe communication bottlenecks. We propose a novel end-to-end hybrid learning framework that allows the edge to selectively query only those hard examples that the cloud can classify correctly. Our framework optimizes over neural architectures and trains edge predictors and routing models so that the overall accuracy remains high while minimizing the overall latency. Training a hybrid learner is difficult since we lack annotations of hard edge-examples. We introduce a novel proxy supervision in this context and show that our method adapts seamlessly and near optimally across different latency regimes. On the ImageNet dataset, our proposed method deployed on a micro-controller unit exhibits $25\%$ reduction in latency compared to cloud-only processing while suffering no excess loss.

Many variants of the Wasserstein distance have been introduced to reduce its original computational burden. In particular the Sliced-Wasserstein distance (SW), which leverages one-dimensional projections for which a closed-form solution of the Wasserstein distance is available, has received a lot of interest. Yet, it is restricted to data living in Euclidean spaces, while the Wasserstein distance has been studied and used recently on manifolds. We focus more specifically on the sphere, for which we define a novel SW discrepancy, which we call spherical Sliced-Wasserstein, making a first step towards defining SW discrepancies on manifolds. Our construction is notably based on closed-form solutions of the Wasserstein distance on the circle, together with a new spherical Radon transform. Along with efficient algorithms and the corresponding implementations, we illustrate its properties in several machine learning use cases where spherical representations of data are at stake: sampling on the sphere, density estimation on real eath data or hyperspherical auto-encoders.

There has been a proliferation of artificial intelligence applications, where model training is key to promising high-quality services for these applications. However, the model training process is both time-intensive and energy-intensive, inevitably affecting the user's demand for application efficiency. Layer freezing, an efficient model training technique, has been proposed to improve training efficiency. Although existing layer freezing methods demonstrate the great potential to reduce model training costs, they still remain shortcomings such as lacking generalizability and compromised accuracy. For instance, existing layer freezing methods either require the freeze configurations to be manually defined before training, which does not apply to different networks, or use heuristic freezing criteria that is hard to guarantee decent accuracy in different scenarios. Therefore, there lacks a generic and smart layer freezing method that can automatically perform ``in-situation'' layer freezing for different networks during training processes. To this end, we propose a generic and efficient training framework (SmartFRZ). The core proposed technique in SmartFRZ is attention-guided layer freezing, which can automatically select the appropriate layers to freeze without compromising accuracy. Experimental results show that SmartFRZ effectively reduces the amount of computation in training and achieves significant training acceleration, and outperforms the state-of-the-art layer freezing approaches.

There is a rising interest in further exploring the zero-shot learning potential of large pre-trained language models (PLMs). A new paradigm called data-generation-based zero-shot learning has achieved impressive success. In this paradigm, the synthesized data from the PLM acts as the carrier of knowledge, which is used to train a task-specific model with orders of magnitude fewer parameters than the PLM, achieving both higher performance and efficiency than prompt-based zero-shot learning methods on PLMs. The main hurdle of this approach is that the synthesized data from PLM usually contains a significant portion of low-quality samples. Fitting on such data will greatly hamper the performance of the task-specific model, making it unreliable for deployment. Previous methods remedy this issue mainly by filtering synthetic data using heuristic metrics(e.g., output confidence), or refining the data with the help of a human expert, which comes with excessive manual tuning or expensive costs. In this paper, we propose a novel noise-robust re-weighting framework SunGen to automatically construct high-quality data for zero-shot classification problems. Our framework features the ability to learn the sample weights indicating data quality without requiring any human annotation. We theoretically and empirically verify the ability of our method to help construct good-quality synthetic datasets. Notably, SunGen-LSTM yields a 9.8% relative improvement than the baseline on average accuracy across eight different established text classification tasks.

Deep neural networks have shown excellent prospects in speech separation tasks. However, obtaining good results while keeping a low model complexity remains challenging in real-world applications. In this paper, we provide a bio-inspired efficient encoder-decoder architecture by mimicking the brain’s top-down attention, called TDANet, with decreased model complexity without sacrificing performance. The top-down attention in TDANet is extracted by the global attention (GA) module and the cascaded local attention (LA) layers. The GA module takes multi-scale acoustic features as input to extract global attention signal, which then modulates features of different scales by direct top-down connections. The LA layers use features of adjacent layers as input to extract the local attention signal, which is used to modulate the lateral input in a top-down manner. On three benchmark datasets, TDANet consistently achieved competitive separation performance to previous state-of-the-art (SOTA) methods with higher efficiency. Specifically, TDANet’s multiply-accumulate operations (MACs) are only 5% of Sepformer, one of the previous SOTA models, and CPU inference time is only 10% of Sepformer. In addition, a large-size version of TDANet obtained SOTA results on three datasets, with MACs still only 10% of Sepformer and the CPU inference time only 24% of Sepformer. Our study suggests that top-down attention can be a more efficient strategy for speech separation.

With the adoption of recent laws ensuring the ``right to be forgotten'', the problem of machine unlearning has become of significant importance. This is particularly the case for graph-structured data, and learning tools specialized for such data, including graph neural networks (GNNs). This work introduces the first known approach for \emph{approximate graph unlearning} with provable theoretical guarantees. The challenges in addressing the problem are two-fold. First, there exist multiple different types of unlearning requests that need to be considered, including node feature, edge and node unlearning. Second, to establish provable performance guarantees, one needs to carefully evaluate the process of feature mixing during propagation. We focus on analyzing Simple Graph Convolutions (SGC) and their generalized PageRank (GPR) extensions, thereby laying the theoretical foundations for unlearning GNNs. Empirical evaluations of six benchmark datasets demonstrate excellent performance/complexity/privacy trade-offs of our approach compared to complete retraining and general methods that do not leverage graph information. For example, unlearning $200$ out of $1208$ training nodes of the Cora dataset only leads to a $0.1\%$ loss in test accuracy, but offers a $4$-fold speed-up compared to complete retraining with a $(\epsilon,\delta)=(1,10^{-4})$ ``privacy cost''. We also exhibit a $12\%$ increase in test accuracy for the same dataset when compared to unlearning methods that do not leverage graph information, with comparable time complexity and the same privacy guarantee.

No existing spherical convolutional neural network (CNN) framework is both computationally scalable and rotationally equivariant. Continuous approaches capture rotational equivariance but are often prohibitively computationally demanding. Discrete approaches offer more favorable computational performance but at the cost of equivariance. We develop a hybrid discrete-continuous (DISCO) group convolution that is simultaneously equivariant and computationally scalable to high-resolution. While our framework can be applied to any compact group, we specialize to the sphere. Our DISCO spherical convolutions exhibit $\text{SO}(3)$ rotational equivariance, where $\text{SO}(n)$ is the special orthogonal group representing rotations in $n$-dimensions. When restricting rotations of the convolution to the quotient space $\text{SO}(3)/\text{SO}(2)$ for further computational enhancements, we recover a form of asymptotic $\text{SO}(3)$ rotational equivariance. Through a sparse tensor implementation we achieve linear scaling in number of pixels on the sphere for both computational cost and memory usage. For 4k spherical images we realize a saving of $10^9$ in computational cost and $10^4$ in memory usage when compared to the most efficient alternative equivariant spherical convolution. We apply the DISCO spherical CNN framework to a number of benchmark dense-prediction problems on the sphere, such as semantic segmentation and depth estimation, on all of which we achieve the state-of-the-art performance.

We introduce the Conditional Independence Regression CovariancE (CIRCE), a measure of conditional independence for multivariate continuous-valued variables. CIRCE applies as a regularizer in settings where we wish to learn neural features $\varphi(X)$ of data $X$ to estimate a target $Y$, while being conditionally independent of a distractor $Z$ given $Y$. Both $Z$ and $Y$ are assumed to be continuous-valued but relatively low dimensional, whereas $X$ and its features may be complex and high dimensional. Relevant settings include domain-invariant learning, fairness, and causal learning. The procedure requires just a single ridge regression from $Y$ to kernelized features of $Z$, which can be done in advance. It is then only necessary to enforce independence of $\varphi(X)$ from residuals of this regression, which is possible with attractive estimation properties and consistency guarantees. By contrast, earlier measures of conditional feature dependence require multiple regressions for each step of feature learning, resulting in more severe bias and variance, and greater computational cost. When sufficiently rich features are used, we establish that CIRCE is zero if and only if $\varphi(X) \perp \!\!\! \perp Z \mid Y$. In experiments, we show superior performance to previous methods on challenging benchmarks, including learning conditionally invariant image features. Code for image data experiments is available at github.com/namratadeka/circe.

Scaling transformers has led to significant breakthroughs in many domains, leading to a paradigm in which larger versions of existing models are trained and released on a periodic basis. New instances of such models are typically trained completely from scratch, despite the fact that they are often just scaled-up versions of their smaller counterparts. How can we use the implicit knowledge in the parameters of smaller, extant models to enable faster training of newer, larger models? This paper describes an approach for accelerating transformer training by learning to grow pretrained transformers, where we learn to linearly map the parameters of the smaller model to initialize the larger model. For tractable learning, we factorize the linear transformation as a composition of (linear) width- and depth-growth operators, and further employ a Kronecker factorization of these growth operators to encode architectural knowledge. Extensive experiments across both language and vision transformers demonstrate that our learned Linear Growth Operator (LiGO) can save up to 50% computational cost of training from scratch, while also consistently outperforming strong baselines that also reuse smaller pretrained models to initialize larger models.

Transformers have quickly shined in the computer vision world since the emergence of Vision Transformers (ViTs). The dominant role of convolutional neural networks (CNNs) seems to be challenged by increasingly effective transformer-based models. Very recently, a couple of advanced convolutional models strike back with large kernels motivated by the local-window attention mechanism, showing appealing performance and efficiency. While one of them, i.e. RepLKNet, impressively manages to scale the kernel size to 31x31 with improved performance, the performance starts to saturate as the kernel size continues growing, compared to the scaling trend of advanced ViTs such as Swin Transformer. In this paper, we explore the possibility of training extreme convolutions larger than 31x31 and test whether the performance gap can be eliminated by strategically enlarging convolutions. This study ends up with a recipe for applying extremely large kernels from the perspective of sparsity, which can smoothly scale up kernels to 61x61 with better performance. Built on this recipe, we propose Sparse Large Kernel Network (SLaK), a pure CNN architecture equipped with sparse factorized 51x51 kernels that can perform on par with or better than state-of-the-art hierarchical Transformers and modern ConvNet architectures like ConvNeXt and RepLKNet, on ImageNet classification as well as a wide range of downstream tasks including semantic segmentation on ADE20K, object detection on PASCAL VOC 2007, and object detection/segmentation on MS COCO. Codes are available at https://github.com/VITA-Group/SLaK.

Predicting the pose of objects from a single image is an important but difficult computer vision problem. Methods that predict a single point estimate do not predict the pose of objects with symmetries well and cannot represent uncertainty. Alternatively, some works predict a distribution over orientations in $\mathrm{SO}(3)$. However, training such models can be computation- and sample-inefficient. Instead, we propose a novel mapping of features from the image domain to the 3D rotation manifold. Our method then leverages $\mathrm{SO}(3)$ equivariant layers, which are more sample efficient, and outputs a distribution over rotations that can be sampled at arbitrary resolution. We demonstrate the effectiveness of our method at object orientation prediction, and achieve state-of-the-art performance on the popular PASCAL3D+ dataset. Moreover, we show that our method can model complex object symmetries, without any modifications to the parameters or loss function. Code is available at \url{https://dmklee.github.io/image2sphere}.

Many point-based 3D detectors adopt point-feature sampling strategies to drop some points for efficient inference. These strategies are typically based on fixed and handcrafted rules, making it difficult to handle complicated scenes. Different from them, we propose a Dynamic Ball Query (DBQ) network to adaptively select a subset of input points according to the input features, and assign the feature transform with a suitable receptive field for each selected point. It can be embedded into some state-of-the-art 3D detectors and trained in an end-to-end manner, which significantly reduces the computational cost. Extensive experiments demonstrate that our method can reduce latency by 30%-100% on KITTI, Waymo, and ONCE datasets. Specifically, the inference speed of our detector can reach 162 FPS on KITTI scene, and 30 FPS on Waymo and ONCE scenes without performance degradation. Due to skipping the redundant points, some evaluation metrics show significant improvements.

While deep learning has achieved great success in various fields, a large amount of memory is necessary to train deep neural networks, which hinders the development of massive state-of-the-art models. The reason is the conventional learning rule, backpropagation, should temporarily store input activations of all the layers in the network. To overcome this, recent studies suggested various memory-efficient implementations of backpropagation. However, those approaches incur computational overhead due to the recomputation of activations, slowing down neural network training. In this work, we propose a new learning rule which significantly reduces memory requirements while closely matching the performance of backpropagation. The algorithm combines auxiliary activation with output activation during forward propagation, while only auxiliary activation is used during backward propagation instead of actual input activation to reduce the amount of data to be temporarily stored. We mathematically show that our learning rule can reliably train the networks whose loss landscape is convex if the auxiliary activation satisfies certain conditions. Based on this observation, we suggest candidates of auxiliary activation that satisfy those conditions. Experimental results confirm that the proposed learning rule achieves competitive performance compared to backpropagation in various models such as ResNet, Transformer, BERT, ViT, and MLP-Mixer.

Dot-product is a central building block in neural networks. However, multiplication ($\texttt{mult}$) in dot-product consumes intensive energy and space costs that challenge deployment on resource-constrained edge devices. In this study, we realize energy-efficient neural networks by exploiting a $\texttt{mult}$-less, sparse dot-product. We first reformulate a dot-product between an integer weight and activation into an equivalent operation comprised of additions followed by bit-shifts ($\texttt{add-shift-add}$). In this formulation, the number of $\texttt{add}$ operations equals the number of bits of the integer weight in binary format. Leveraging this observation, we propose Bit-Pruning, which removes unnecessary bits in each weight value during training to reduce the energy consumption of $\texttt{add-shift-add}$. Bit-Pruning can be seen as soft Weight-Pruning as it prunes bits, not the whole weight element. In extensive experiments, we demonstrate that sparse $\texttt{mult}$-less networks trained with Bit-Pruning show a better accuracy-energy trade-off than sparse $\texttt{mult}$ networks trained with Weight-Pruning.

Deep Ensembles (DE) are a prominent approach for achieving excellent performance on key metrics such as accuracy, calibration, uncertainty estimation, and out-of-distribution detection. However, hardware limitations of real-world systems constrain to smaller ensembles and lower-capacity networks, significantly deteriorating their performance and properties. We introduce Packed-Ensembles (PE), a strategy to design and train lightweight structured ensembles by carefully modulating the dimension of their encoding space. We leverage grouped convolutions to parallelize the ensemble into a single shared backbone and forward pass to improve training and inference speeds. PE is designed to operate within the memory limits of a standard neural network. Our extensive research indicates that PE accurately preserves the properties of DE, such as diversity, and performs equally well in terms of accuracy, calibration, out-of-distribution detection, and robustness to distribution shift. We make our code available at https://github.com/ENSTA-U2IS/torch-uncertainty.

Parameter-efficient fine-tuning aims to achieve comparable performances of fine-tuning with much fewer trainable parameters. Recently, various tuning strategies (e.g., Adapters, Prefix Tuning, BitFit, and LoRA) have been proposed. However, their designs are hand-crafted separately, and it remains unclear whether certain design patterns exist for parameter-efficient fine-tuning. Thus, we present a parameter-efficient fine-tuning design paradigm and discover design patterns that are applicable to different experimental settings. Instead of focusing on designing another individual tuning strategy, we introduce parameter-efficient fine-tuning design spaces that parameterize tuning structures and tuning strategies. Specifically, any design space is characterized by four components: layer grouping, trainable parameter allocation, tunable groups, and strategy assignment. Our comprehensive empirical study leads to the discovery of design patterns: (i) grouping layers in a spindle pattern, (ii) uniformly allocating the number of trainable parameters to layers, (ii) tuning all the groups, and (iv) tuning different groups with proper strategies. Our discovered design patterns result in new parameter-efficient fine-tuning methods. Experiments show that these methods consistently outperform investigated parameter-efficient fine-tuning strategies across different backbone models and different tasks in natural language processing.

We present a computationally efficient framework to model a wide range of population structures with high order correlations and a large number of neurons. Our method is based on a special type of Bayesian network that has linear inference time and is founded upon the concept of contextual independence. Moreover, we use an efficient architecture learning method for network selection to model large neural populations even with a small amount of data. Our framework is both fast and accurate in approximating neural population structures. Furthermore, our approach enables us to reliably quantify higher order neural correlations. We test our method on simulated neural populations commonly used to generate higher order correlations, as well as on publicly available large-scale neural recordings from the Allen Brain Observatory. Our approach significantly outperforms other models both in terms of statistical measures and alignment with experimental evidence.

In this paper, we propose Block and Subword-Scaling Floating-Point (BSFP), a non-uniform quantization scheme for the skewed and non-uniform distribution of weight vectors in neural networks. By quantizing each weight vector as the superposition of multiple subword vectors (in two's complement) with scaling factors (in Low-bit Floating-Point, LBFP), BSFP can effectively fit the distribution of weight vectors while maintaining high computation efficiency. Furthermore, we present a grid search-based MSE-optimal quantization flow and a scaled serial processing engine to complete the quantization pipeline and the infrastructure. The experimental results on the ImageNet classification task show that our proposed method outperforms state-of-the-art Microsoft Floating Point (MSFP) by up to 20.56% top-1 accuracy at the same weight precision and reduces up to 10.3% model size. Furthermore, BSFP outperforms MSFP by up to 2.0$\times$ computing throughput and up to 5.3$\times$ energy efficiency under the same silicon area budget.

Deep generative models parametrized up to a normalizing constant (e.g. energy-based models) are difficult to train by maximizing the likelihood of the data because the likelihood and/or gradients thereof cannot be explicitly or efficiently written down. Score matching is a training method, whereby instead of fitting the likelihood $\log p(x)$ for the training data, we instead fit the score function $\nabla_x \log p(x)$ --- obviating the need to evaluate the partition function. Though this estimator is known to be consistent, its unclear whether (and when) its statistical efficiency is comparable to that of maximum likelihood --- which is known to be (asymptotically) optimal. We initiate this line of inquiry in this paper, and show a tight connection between statistical efficiency of score matching and the isoperimetric properties of the distribution being estimated --- i.e. the Poincar\'e, log-Sobolev and isoperimetric constant --- quantities which govern the mixing time of Markov processes like Langevin dynamics. Roughly, we show that the score matching estimator is statistically comparable to the maximum likelihood when the distribution has a small isoperimetric constant. Conversely, if the distribution has a large isoperimetric constant --- even for simple families of distributions like exponential families with rich enough sufficient statistics --- score matching will be substantially less efficient than maximum likelihood. We suitably formalize these results both in the finite sample regime, and in the asymptotic regime. Finally, we identify a direct parallel in the discrete setting, where we connect the statistical properties of pseudolikelihood estimation with approximate tensorization of entropy and the Glauber dynamics.