We propose Adaptive Compressed Gradient Descent (AdaCGD) - a novel optimization algorithm for communication-efficient training of supervised machine learning models with adaptive compression level. Our approach is inspired by the recently proposed three point compressor (3PC) framework of Richtarik et al. (2022), which includes error feedback (EF21), lazily aggregated gradient (LAG), and their combination as special cases, and offers the current state-of-the-art rates for these methods under weak assumptions. While the above mechanisms offer a fixed compression level, or adapt between two extremes only, our proposal is to perform a much finer adaptation. In particular, we allow the user to choose any number of arbitrarily chosen contractive compression mechanisms, such as Top-K sparsification with a user-defined selection of sparsification levels K, or quantization with a user-defined selection of quantization levels, or their combination. AdaCGD chooses the appropriate compressor and compression level adaptively during the optimization process. Besides i) proposing a theoretically-grounded multi-adaptive communication compression mechanism, we further ii) extend the 3PC framework to bidirectional compression, i.e., we allow the server to compress as well, and iii) provide sharp convergence bounds in the strongly convex, convex and nonconvex settings. The convex regime results are new even for several key special cases of our general mechanism, including 3PC and EF21. In all regimes, our rates are superior compared to all existing adaptive compression methods.
DNN-based speaker verification (SV) models demonstrate significant performance at relatively high computation costs. Model compression can be applied to reduce the model size for lower resource consumption. The present study exploits weight quantization to compress two widely-used SV models, namely ECAPA-TDNN and ResNet. Experimental results on VoxCeleb show that weight quantization is effective for compressing SV models. The model size can be reduced multiple times without noticeable degradation in performance. Compression of ResNet shows more robust results than ECAPA-TDNN with lower-bitwidth quantization. Analysis of the layer weights suggests that the smooth weight distribution of ResNet may be related to its better robustness. The generalization ability of the quantized model is validated via a language-mismatched SV task. Furthermore, analysis by information probing reveals that the quantized models can retain most of the speaker-relevant knowledge learned by the original models.
Efficient document retrieval heavily relies on the technique of semantic hashing, which learns a binary code for every document and employs Hamming distance to evaluate document distances. However, existing semantic hashing methods are mostly established on outdated TFIDF features, which obviously do not contain lots of important semantic information about documents. Furthermore, the Hamming distance can only be equal to one of several integer values, significantly limiting its representational ability for document distances. To address these issues, in this paper, we propose to leverage BERT embeddings to perform efficient retrieval based on the product quantization technique, which will assign for every document a real-valued codeword from the codebook, instead of a binary code as in semantic hashing. Specifically, we first transform the original BERT embeddings via a learnable mapping and feed the transformed embedding into a probabilistic product quantization module to output the assigned codeword. The refining and quantizing modules can be optimized in an end-to-end manner by minimizing the probabilistic contrastive loss. A mutual information maximization based method is further proposed to improve the representativeness of codewords, so that documents can be quantized more accurately. Extensive experiments conducted on three benchmarks demonstrate that our proposed method significantly outperforms current state-of-the-art baselines.
Transformer-based architectures like BERT have achieved great success in a wide range of Natural Language tasks. Despite their decent performance, the models still have numerous parameters and high computational complexity, impeding their deployment in resource-constrained environments. Post-Training Quantization (PTQ), which enables low-bit computations without extra training, could be a promising tool. In this work, we conduct an empirical evaluation of three PTQ methods on BERT-Base and BERT-Large: Linear Quantization (LQ), Analytical Clipping for Integer Quantization (ACIQ), and Outlier Channel Splitting (OCS). OCS theoretically surpasses the others in minimizing the Mean Square quantization Error and avoiding distorting the weights' outliers. That is consistent with the evaluation results of most language tasks of GLUE benchmark and a reading comprehension task, SQuAD. Moreover, low-bit quantized BERT models could outperform the corresponding 32-bit baselines on several small language tasks, which we attribute to the alleviation of over-parameterization. We further explore the limit of quantization bit and show that OCS could quantize BERT-Base and BERT-Large to 3-bits and retain 98% and 96% of the performance on the GLUE benchmark accordingly. Moreover, we conduct quantization on the whole BERT family, i.e., BERT models in different configurations, and comprehensively evaluate their performance on the GLUE benchmark and SQuAD, hoping to provide valuable guidelines for their deployment in various computation environments.
Limited computational budgets often prevent transformers from being used in production and from having their high accuracy utilized. A knowledge distillation approach addresses the computational efficiency by self-distilling BERT into a smaller transformer representation having fewer layers and smaller internal embedding. However, the performance of these models drops as we reduce the number of layers, notably in advanced NLP tasks such as span question answering. In addition, a separate model must be trained for each inference scenario with its distinct computational budget. Dynamic-TinyBERT tackles both limitations by partially implementing the Length Adaptive Transformer (LAT) technique onto TinyBERT, achieving x3 speedup over BERT-base with minimal accuracy loss. In this work, we expand the Dynamic-TinyBERT approach to generate a much more highly efficient model. We use MiniLM distillation jointly with the LAT method, and we further enhance the efficiency by applying low-bit quantization. Our quantized length-adaptive MiniLM model (QuaLA-MiniLM) is trained only once, dynamically fits any inference scenario, and achieves an accuracy-efficiency trade-off superior to any other efficient approaches per any computational budget on the SQuAD1.1 dataset (up to x8.8 speedup with <1% accuracy loss). The code to reproduce this work is publicly available on Github.
Federated clustering (FC) is an unsupervised learning problem that arises in a number of practical applications, including personalized recommender and healthcare systems. With the adoption of recent laws ensuring the "right to be forgotten", the problem of machine unlearning for FC methods has become of significant importance. We introduce, for the first time, the problem of machine unlearning for FC, and propose an efficient unlearning mechanism for a customized secure FC framework. Our FC framework utilizes special initialization procedures that we show are well-suited for unlearning. To protect client data privacy, we develop the secure compressed multiset aggregation (SCMA) framework that addresses sparse secure federated learning (FL) problems encountered during clustering as well as more general problems. To simultaneously facilitate low communication complexity and secret sharing protocols, we integrate Reed-Solomon encoding with special evaluation points into our SCMA pipeline, and prove that the client communication cost is logarithmic in the vector dimension. Additionally, to demonstrate the benefits of our unlearning mechanism over complete retraining, we provide a theoretical analysis for the unlearning performance of our approach. Simulation results show that the new FC framework exhibits superior clustering performance compared to previously reported FC baselines when the cluster sizes are highly imbalanced. Compared to completely retraining K-means++ locally and globally for each removal request, our unlearning procedure offers an average speed-up of roughly 84x across seven datasets. Our implementation for the proposed method is available at https://github.com/thupchnsky/mufc.
Data-parallel distributed training of deep neural networks (DNN) has gained very widespread adoption, but can still experience communication bottlenecks. To address this issue, entire families of compression mechanisms have been developed, including quantization, sparsification, and low-rank approximation, some of which are seeing significant practical adoption. Despite this progress, almost all known compression schemes apply compression uniformly across DNN layers, although layers are heterogeneous in terms of parameter count and their impact on model accuracy. In this work, we provide a general framework for adapting the degree of compression across the model's layers dynamically during training, improving the overall compression, while leading to substantial speedups, without sacrificing accuracy. Our framework, called L-GreCo, is based on an adaptive algorithm, which automatically picks the optimal compression parameters for model layers guaranteeing the best compression ratio while satisfying an error constraint. Extensive experiments over image classification and language modeling tasks shows that L-GreCo is effective across all existing families of compression methods, and achieves up to 2.5$\times$ training speedup and up to 5$\times$ compression improvement over efficient implementations of existing approaches, while recovering full accuracy. Moreover, L-GreCo is complementary to existing adaptive algorithms, improving their compression ratio by 50% and practical throughput by 66%.
Generative Pre-trained Transformer models, known as GPT or OPT, set themselves apart through breakthrough performance across complex language modelling tasks, but also by their extremely high computational and storage costs. Specifically, due to their massive size, even inference for large, highly-accurate GPT models may require multiple performant GPUs, which limits the usability of such models. While there is emerging work on relieving this pressure via model compression, the applicability and performance of existing compression techniques is limited by the scale and complexity of GPT models. In this paper, we address this challenge, and propose GPTQ, a new one-shot weight quantization method based on approximate second-order information, that is both highly-accurate and highly-efficient. Specifically, GPTQ can quantize GPT models with 175 billion parameters in approximately four GPU hours, reducing the bitwidth down to 3 or 4 bits per weight, with negligible accuracy degradation relative to the uncompressed baseline. Our method more than doubles the compression gains relative to previously-proposed one-shot quantization methods, preserving accuracy, allowing us for the first time to execute an 175 billion-parameter model inside a single GPU for generative inference. Moreover, we also show that our method can still provide reasonable accuracy in the extreme quantization regime, in which weights are quantized to 2-bit or even ternary quantization levels. We show experimentally that these improvements can be leveraged for end-to-end inference speedups over FP16, of around 3.25x when using high-end GPUs (NVIDIA A100) and 4.5x when using more cost-effective ones (NVIDIA A6000). The implementation is available at https://github.com/IST-DASLab/gptq.
Spiking neural networks (SNNs) have closer dynamics to the brain than current deep neural networks. Their low power consumption and sample efficiency make these networks interesting. Recently, several deep convolutional spiking neural networks have been proposed. These networks aim to increase biological plausibility while creating powerful tools to be applied to machine learning tasks. Here, we suggest a network structure based on previous work to improve network runtime and accuracy. Improvements to the network include reducing training iterations to only once, effectively using principal component analysis (PCA) dimension reduction, weight quantization, timed outputs for classification, and better hyperparameter tuning. Furthermore, the preprocessing step is changed to allow the processing of colored images instead of only black and white to improve accuracy. The proposed structure fractionalizes runtime and introduces an efficient approach to deep convolutional SNNs.
Benefit from flexible network designs and end-to-end joint optimization approach, learned image compression (LIC) has demonstrated excellent coding performance and practical feasibility in recent years. However, existing compression models suffer from serious multi-generation loss, which always occurs during image editing and transcoding. During the process of repeatedly encoding and decoding, the quality of the image will rapidly degrade, resulting in various types of distortion, which significantly limits the practical application of LIC. In this paper, a thorough analysis is carried out to determine the source of generative loss in successive image compression (SIC). We point out and solve the quantization drift problem that affects SIC, reversibility loss function as well as channel relaxation method are proposed to further reduce the generation loss. Experiments show that by using our proposed solutions, LIC can achieve comparable performance to the first compression of BPG even after 50 times reencoding without any change of the network structure.
Data quantization is an effective method to accelerate neural network training and reduce power consumption. However, it is challenging to perform low-bit quantized training: the conventional equal-precision quantization will lead to either high accuracy loss or limited bit-width reduction, while existing mixed-precision methods offer high compression potential but failed to perform accurate and efficient bit-width assignment. In this work, we propose DYNASTY, a block-wise dynamic-precision neural network training framework. DYNASTY provides accurate data sensitivity information through fast online analytics, and maintains stable training convergence with an adaptive bit-width map generator. Network training experiments on CIFAR-100 and ImageNet dataset are carried out, and compared to 8-bit quantization baseline, DYNASTY brings up to $5.1\times$ speedup and $4.7\times$ energy consumption reduction with no accuracy drop and negligible hardware overhead.
Multiresolution analysis has applications across many disciplines in the study of complex systems and their dynamics. Financial markets are among the most complex entities in our environment, yet mainstream quantitative models operate at predetermined scale, rely on linear correlation measures, and struggle to recognize non-linear or causal structures. In this paper, we combine neural networks known to capture non-linear associations with a multiscale decomposition to facilitate a better understanding of financial market data substructures. Quantization keeps our decompositions calibrated to market at every scale. We illustrate our approach in the context of seven use cases.
Knowledge distillation which learns a lightweight student model by distilling knowledge from a cumbersome teacher model is an attractive approach for learning compact deep neural networks (DNNs). Recent works further improve student network performance by leveraging multiple teacher networks. However, most of the existing knowledge distillation-based multi-teacher methods use separately pretrained teachers. This limits the collaborative learning between teachers and the mutual learning between teachers and student. Network quantization is another attractive approach for learning compact DNNs. However, most existing network quantization methods are developed and evaluated without considering multi-teacher support to enhance the performance of quantized student model. In this paper, we propose a novel framework that leverages both multi-teacher knowledge distillation and network quantization for learning low bit-width DNNs. The proposed method encourages both collaborative learning between quantized teachers and mutual learning between quantized teachers and quantized student. During learning process, at corresponding layers, knowledge from teachers will form an importance-aware shared knowledge which will be used as input for teachers at subsequent layers and also be used to guide student. Our experimental results on CIFAR100 and ImageNet datasets show that the compact quantized student models trained with our method achieve competitive results compared to other state-of-the-art methods, and in some cases, indeed surpass the full precision models.
Leveraging shared learning through Massively Multilingual Models, state-of-the-art machine translation models are often able to adapt to the paucity of data for low-resource languages. However, this performance comes at the cost of significantly bloated models which are not practically deployable. Knowledge Distillation is one popular technique to develop competitive, lightweight models: In this work, we first evaluate its use to compress MT models focusing on languages with extremely limited training data. Through our analysis across 8 languages, we find that the variance in the performance of the distilled models due to their dependence on priors including the amount of synthetic data used for distillation, the student architecture, training hyperparameters and confidence of the teacher models, makes distillation a brittle compression mechanism. To mitigate this, we explore the use of post-training quantization for the compression of these models. Here, we find that while distillation provides gains across some low-resource languages, quantization provides more consistent performance trends for the entire range of languages, especially the lowest-resource languages in our target set.
Search engines and recommendation systems are built to efficiently display relevant information from those massive amounts of candidates. Typically a three-stage mechanism is employed in those systems: (i) a small collection of items are first retrieved by (e.g.,) approximate near neighbor search algorithms; (ii) then a collection of constraints are applied on the retrieved items; (iii) a fine-grained ranking neural network is employed to determine the final recommendation. We observe a major defect of the original three-stage pipeline: Although we only target to retrieve $k$ vectors in the final recommendation, we have to preset a sufficiently large $s$ ($s > k$) for each query, and ``hope'' the number of survived vectors after the filtering is not smaller than $k$. That is, at least $k$ vectors in the $s$ similar candidates satisfy the query constraints. In this paper, we investigate this constrained similarity search problem and attempt to merge the similarity search stage and the filtering stage into one single search operation. We introduce AIRSHIP, a system that integrates a user-defined function filtering into the similarity search framework. The proposed system does not need to build extra indices nor require prior knowledge of the query constraints. We propose three optimization strategies: (1) starting point selection, (2) multi-direction search, and (3) biased priority queue selection. Experimental evaluations on both synthetic and real data confirm the effectiveness of the proposed AIRSHIP algorithm. We focus on constrained graph-based approximate near neighbor (ANN) search in this study, in part because graph-based ANN is known to achieve excellent performance. We believe it is also possible to develop constrained hashing-based ANN or constrained quantization-based ANN.
Contemporary lossy image and video coding standards rely on transform coding, the process through which pixels are mapped to an alternative representation to facilitate efficient data compression. Despite impressive performance of end-to-end optimized compression with deep neural networks, the high computational and space demands of these models has prevented them from superseding the relatively simple transform coding found in conventional video codecs. In this study, we propose learned transforms and entropy coding that may either serve as (non)linear drop-in replacements, or enhancements for linear transforms in existing codecs. These transforms can be multi-rate, allowing a single model to operate along the entire rate-distortion curve. To demonstrate the utility of our framework, we augmented the DCT with learned quantization matrices and adaptive entropy coding to compress intra-frame AV1 block prediction residuals. We report substantial BD-rate and perceptual quality improvements over more complex nonlinear transforms at a fraction of the computational cost.
Enabling low precision implementations of deep learning models, without considerable performance degradation, is necessary in resource and latency constrained settings. Moreover, exploiting the differences in sensitivity to quantization across layers can allow mixed precision implementations to achieve a considerably better computation performance trade-off. However, backpropagating through the quantization operation requires introducing gradient approximations, and choosing which layers to quantize is challenging for modern architectures due to the large search space. In this work, we present a constrained learning approach to quantization aware training. We formulate low precision supervised learning as a constrained optimization problem, and show that despite its non-convexity, the resulting problem is strongly dual and does away with gradient estimations. Furthermore, we show that dual variables indicate the sensitivity of the objective with respect to constraint perturbations. We demonstrate that the proposed approach exhibits competitive performance in image classification tasks, and leverage the sensitivity result to apply layer selective quantization based on the value of dual variables, leading to considerable performance improvements.
We present a fully on-device streaming Speech2Speech conversion model that normalizes a given input speech directly to synthesized output speech. Deploying such a model on mobile devices pose significant challenges in terms of memory footprint and computation requirements. We present a streaming-based approach to produce an acceptable delay, with minimal loss in speech conversion quality, when compared to a reference state of the art non-streaming approach. Our method consists of first streaming the encoder in real time while the speaker is speaking. Then, as soon as the speaker stops speaking, we run the spectrogram decoder in streaming mode along the side of a streaming vocoder to generate output speech. To achieve an acceptable delay-quality trade-off, we propose a novel hybrid approach for look-ahead in the encoder which combines a look-ahead feature stacker with a look-ahead self-attention. We show that our streaming approach is almost 2x faster than real time on the Pixel4 CPU.
We propose a novel method for training a conditional generative adversarial network (CGAN) without the use of training data, called zero-shot learning of a CGAN (ZS-CGAN). Zero-shot learning of a conditional generator only needs a pre-trained discriminative (classification) model and does not need any training data. In particular, the conditional generator is trained to produce labeled synthetic samples whose characteristics mimic the original training data by using the statistics stored in the batch normalization layers of the pre-trained model. We show the usefulness of ZS-CGAN in data-free quantization of deep neural networks. We achieved the state-of-the-art data-free network quantization of the ResNet and MobileNet classification models trained on the ImageNet dataset. Data-free quantization using ZS-CGAN showed a minimal loss in accuracy compared to that obtained by conventional data-dependent quantization.
We present a smoothly broken power law functional form (referred to by us as a Broken Neural Scaling Law (BNSL)) that accurately models and extrapolates the scaling behaviors of deep neural networks (i.e. how the evaluation metric of interest varies as the amount of compute used for training, number of model parameters, training dataset size, model input size, number of training steps, or upstream performance varies) for various architectures and for each of various tasks within a large and diverse set of upstream and downstream tasks, in zero-shot, prompted, and fine-tuned settings. This set includes large-scale vision, language, audio, video, diffusion, generative modeling, multimodal learning, contrastive learning, AI alignment, robotics, out-of-distribution (OOD) generalization, continual learning, transfer learning, uncertainty estimation / calibration, out-of-distribution detection, adversarial robustness, distillation, sparsity, retrieval, quantization, pruning, molecules, computer programming/coding, math word problems, "emergent" "phase transitions / changes", arithmetic, unsupervised/self-supervised learning, and reinforcement learning (single agent and multi-agent). When compared to other functional forms for neural scaling behavior, this functional form yields extrapolations of scaling behavior that are considerably more accurate on this set. Moreover, this functional form accurately models and extrapolates scaling behavior that other functional forms are incapable of expressing such as the non-monotonic transitions present in the scaling behavior of phenomena such as double descent and the delayed, sharp inflection points present in the scaling behavior of tasks such as arithmetic. Lastly, we use this functional form to glean insights about the limit of the predictability of scaling behavior. Code is available at https://github.com/ethancaballero/broken_neural_scaling_laws