Automatic grading and feedback have been long studied using traditional machine learning and deep learning techniques using language models. With the recent accessibility to high performing large language models (LLMs) like LLaMA-2, there is an opportunity to investigate the use of these LLMs for automatic grading and feedback generation. Despite the increase in performance, LLMs require significant computational resources for fine-tuning and additional specific adjustments to enhance their performance for such tasks. To address these issues, Parameter Efficient Fine-tuning (PEFT) methods, such as LoRA and QLoRA, have been adopted to decrease memory and computational requirements in model fine-tuning. This paper explores the efficacy of PEFT-based quantized models, employing classification or regression head, to fine-tune LLMs for automatically assigning continuous numerical grades to short answers and essays, as well as generating corresponding feedback. We conducted experiments on both proprietary and open-source datasets for our tasks. The results show that prediction of grade scores via finetuned LLMs are highly accurate, achieving less than 3% error in grade percentage on average. For providing graded feedback fine-tuned 4-bit quantized LLaMA-2 13B models outperform competitive base models and achieve high similarity with subject matter expert feedback in terms of high BLEU and ROUGE scores and qualitatively in terms of feedback. The findings from this study provide important insights into the impacts of the emerging capabilities of using quantization approaches to fine-tune LLMs for various downstream tasks, such as automatic short answer scoring and feedback generation at comparatively lower costs and latency.
In this paper, we investigate self-supervised pre-training methods for document text recognition. Nowadays, large unlabeled datasets can be collected for many research tasks, including text recognition, but it is costly to annotate them. Therefore, methods utilizing unlabeled data are researched. We study self-supervised pre-training methods based on masked label prediction using three different approaches -- Feature Quantization, VQ-VAE, and Post-Quantized AE. We also investigate joint-embedding approaches with VICReg and NT-Xent objectives, for which we propose an image shifting technique to prevent model collapse where it relies solely on positional encoding while completely ignoring the input image. We perform our experiments on historical handwritten (Bentham) and historical printed datasets mainly to investigate the benefits of the self-supervised pre-training techniques with different amounts of annotated target domain data. We use transfer learning as strong baselines. The evaluation shows that the self-supervised pre-training on data from the target domain is very effective, but it struggles to outperform transfer learning from closely related domains. This paper is one of the first researches exploring self-supervised pre-training in document text recognition, and we believe that it will become a cornerstone for future research in this area. We made our implementation of the investigated methods publicly available at https://github.com/DCGM/pero-pretraining.
Vision Transformers (ViTs) have recently garnered considerable attention, emerging as a promising alternative to convolutional neural networks (CNNs) in several vision-related applications. However, their large model sizes and high computational and memory demands hinder deployment, especially on resource-constrained devices. This underscores the necessity of algorithm-hardware co-design specific to ViTs, aiming to optimize their performance by tailoring both the algorithmic structure and the underlying hardware accelerator to each other's strengths. Model quantization, by converting high-precision numbers to lower-precision, reduces the computational demands and memory needs of ViTs, allowing the creation of hardware specifically optimized for these quantized algorithms, boosting efficiency. This article provides a comprehensive survey of ViTs quantization and its hardware acceleration. We first delve into the unique architectural attributes of ViTs and their runtime characteristics. Subsequently, we examine the fundamental principles of model quantization, followed by a comparative analysis of the state-of-the-art quantization techniques for ViTs. Additionally, we explore the hardware acceleration of quantized ViTs, highlighting the importance of hardware-friendly algorithm design. In conclusion, this article will discuss ongoing challenges and future research paths. We consistently maintain the related open-source materials at https://github.com/DD-DuDa/awesome-vit-quantization-acceleration.
Model size and inference speed at deployment time, are major challenges in many deep learning applications. A promising strategy to overcome these challenges is quantization. However, a straightforward uniform quantization to very low precision can result in significant accuracy loss. Mixed-precision quantization, based on the idea that certain parts of the network can accommodate lower precision without compromising performance compared to other parts, offers a potential solution. In this work, we present High Granularity Quantization (HGQ), an innovative quantization-aware training method that could fine-tune the per-weight and per-activation precision by making them optimizable through gradient descent. This approach enables ultra-low latency and low power neural networks on hardware capable of performing arithmetic operations with an arbitrary number of bits, such as FPGAs and ASICs. We demonstrate that HGQ can outperform existing methods by a substantial margin, achieving resource reduction by up to a factor of 20 and latency improvement by a factor of 5 while preserving accuracy.
Recent studies introduced effective compression techniques for Large Language Models (LLMs) via post-training quantization or low-bit weight representation. Although quantized weights offer storage efficiency and allow for faster inference, existing works have indicated that quantization might compromise performance and exacerbate biases in LLMs. This study investigates the confidence and calibration of quantized models, considering factors such as language model type and scale as contributors to quantization loss. Firstly, we reveal that quantization with GPTQ to 4-bit results in a decrease in confidence regarding true labels, with varying impacts observed among different language models. Secondly, we observe fluctuations in the impact on confidence across different scales. Finally, we propose an explanation for quantization loss based on confidence levels, indicating that quantization disproportionately affects samples where the full model exhibited low confidence levels in the first place.
Weight quantization is used to deploy high-performance deep learning models on resource-limited hardware, enabling the use of low-precision integers for storage and computation. Spiking neural networks (SNNs) share the goal of enhancing efficiency, but adopt an 'event-driven' approach to reduce the power consumption of neural network inference. While extensive research has focused on weight quantization, quantization-aware training (QAT), and their application to SNNs, the precision reduction of state variables during training has been largely overlooked, potentially diminishing inference performance. This paper introduces two QAT schemes for stateful neurons: (i) a uniform quantization strategy, an established method for weight quantization, and (ii) threshold-centered quantization, which allocates exponentially more quantization levels near the firing threshold. Our results show that increasing the density of quantization levels around the firing threshold improves accuracy across several benchmark datasets. We provide an ablation analysis of the effects of weight and state quantization, both individually and combined, and how they impact models. Our comprehensive empirical evaluation includes full precision, 8-bit, 4-bit, and 2-bit quantized SNNs, using QAT, stateful QAT (SQUAT), and post-training quantization methods. The findings indicate that the combination of QAT and SQUAT enhance performance the most, but given the choice of one or the other, QAT improves performance by the larger degree. These trends are consistent all datasets. Our methods have been made available in our Python library snnTorch: https://github.com/jeshraghian/snntorch.
Quantization-aware training (QAT) simulates a quantization process during training to lower bit-precision of weights/activations. It learns quantized weights indirectly by updating latent weights, i.e., full-precision inputs to a quantizer, using gradient-based optimizers. We claim that coupling a user-defined learning rate (LR) with these optimizers is sub-optimal for QAT. Quantized weights transit discrete levels of a quantizer, only if corresponding latent weights pass transition points, where the quantizer changes discrete states. This suggests that the changes of quantized weights are affected by both the LR for latent weights and their distributions. It is thus difficult to control the degree of changes for quantized weights by scheduling the LR manually. We conjecture that the degree of parameter changes in QAT is related to the number of quantized weights transiting discrete levels. Based on this, we introduce a transition rate (TR) scheduling technique that controls the number of transitions of quantized weights explicitly. Instead of scheduling a LR for latent weights, we schedule a target TR of quantized weights, and update the latent weights with a novel transition-adaptive LR (TALR), enabling considering the degree of changes for the quantized weights during QAT. Experimental results demonstrate the effectiveness of our approach on standard benchmarks.
In this study, we propose the first hardware implementation of a context-based recurrent spiking neural network (RSNN) emphasizing on integrating dual information streams within the neocortical pyramidal neurons specifically Context- Dependent Leaky Integrate and Fire (CLIF) neuron models, essential element in RSNN. We present a quantized version of the CLIF neuron (qCLIF), developed through a hardware-software codesign approach utilizing the sparse activity of RSNN. Implemented in a 45nm technology node, the qCLIF is compact (900um^2) and achieves a high accuracy of 90% despite 8 bit quantization on DVS gesture classification dataset. Our analysis spans a network configuration from 10 to 200 qCLIF neurons, supporting up to 82k synapses within a 1.86 mm^2 footprint, demonstrating scalability and efficiency
Large-scale deep neural networks (DNNs) have achieved remarkable success in many application scenarios. However, high computational complexity and energy costs of modern DNNs make their deployment on edge devices challenging. Model quantization is a common approach to deal with deployment constraints, but searching for optimized bit-widths can be challenging. In this work, we present Adaptive Bit-Width Quantization Aware Training (AdaQAT), a learning-based method that automatically optimizes weight and activation signal bit-widths during training for more efficient DNN inference. We use relaxed real-valued bit-widths that are updated using a gradient descent rule, but are otherwise discretized for all quantization operations. The result is a simple and flexible QAT approach for mixed-precision uniform quantization problems. Compared to other methods that are generally designed to be run on a pretrained network, AdaQAT works well in both training from scratch and fine-tuning scenarios.Initial results on the CIFAR-10 and ImageNet datasets using ResNet20 and ResNet18 models, respectively, indicate that our method is competitive with other state-of-the-art mixed-precision quantization approaches.
This paper investigates three different parameterizations of asymmetric uniform quantization for quantization-aware training: (1) scale and offset, (2) minimum and maximum, and (3) beta and gamma. We perform a comprehensive comparative analysis of these parameterizations' influence on quantization-aware training, using both controlled experiments and real-world large language models. Our particular focus is on their changing behavior in response to critical training hyperparameters, bit width and learning rate. Based on our investigation, we propose best practices to stabilize and accelerate quantization-aware training with learnable asymmetric quantization ranges.
Multimodal recommendation aims to recommend user-preferred candidates based on her/his historically interacted items and associated multimodal information. Previous studies commonly employ an embed-and-retrieve paradigm: learning user and item representations in the same embedding space, then retrieving similar candidate items for a user via embedding inner product. However, this paradigm suffers from inference cost, interaction modeling, and false-negative issues. Toward this end, we propose a new MMGRec model to introduce a generative paradigm into multimodal recommendation. Specifically, we first devise a hierarchical quantization method Graph RQ-VAE to assign Rec-ID for each item from its multimodal and CF information. Consisting of a tuple of semantically meaningful tokens, Rec-ID serves as the unique identifier of each item. Afterward, we train a Transformer-based recommender to generate the Rec-IDs of user-preferred items based on historical interaction sequences. The generative paradigm is qualified since this model systematically predicts the tuple of tokens identifying the recommended item in an autoregressive manner. Moreover, a relation-aware self-attention mechanism is devised for the Transformer to handle non-sequential interaction sequences, which explores the element pairwise relation to replace absolute positional encoding. Extensive experiments evaluate MMGRec's effectiveness compared with state-of-the-art methods.
Large language models (LLMs) are rapidly emerging in Artificial Intelligence (AI) applications, especially in the fields of natural language processing and generative AI. Not limited to text generation applications, these models inherently possess the opportunity to leverage prompt engineering, where the inputs of such models can be appropriately structured to articulate a model's purpose explicitly. A prominent example of this is intent-based networking, an emerging approach for automating and maintaining network operations and management. This paper presents semantic routing to achieve enhanced performance in LLM-assisted intent-based management and orchestration of 5G core networks. This work establishes an end-to-end intent extraction framework and presents a diverse dataset of sample user intents accompanied by a thorough analysis of the effects of encoders and quantization on overall system performance. The results show that using a semantic router improves the accuracy and efficiency of the LLM deployment compared to stand-alone LLMs with prompting architectures.
The recent advancement in deep learning (DL) for automatic modulation classification (AMC) of wireless signals has encouraged numerous possible applications on resource-constrained edge devices. However, developing optimized DL models suitable for edge applications of wireless communications is yet to be studied in depth. In this work, we perform a thorough investigation of optimized convolutional neural networks (CNNs) developed for AMC using the three most commonly used model optimization techniques: a) pruning, b) quantization, and c) knowledge distillation. Furthermore, we have proposed optimized models with the combinations of these techniques to fuse the complementary optimization benefits. The performances of all the proposed methods are evaluated in terms of sparsity, storage compression for network parameters, and the effect on classification accuracy with a reduction in parameters. The experimental results show that the proposed individual and combined optimization techniques are highly effective for developing models with significantly less complexity while maintaining or even improving classification performance compared to the benchmark CNNs.
Embedding-based retrieval serves as a dominant approach to candidate item matching for industrial recommender systems. With the success of generative AI, generative retrieval has recently emerged as a new retrieval paradigm for recommendation, which casts item retrieval as a generation problem. Its model consists of two stages: semantic tokenization and autoregressive generation. The first stage involves item tokenization that constructs discrete semantic tokens to index items, while the second stage autoregressively generates semantic tokens of candidate items. Therefore, semantic tokenization serves as a crucial preliminary step for training generative recommendation models. Existing research usually employs a vector quantizier with reconstruction loss (e.g., RQ-VAE) to obtain semantic tokens of items, but this method fails to capture the essential neighborhood relationships that are vital for effective item modeling in recommender systems. In this paper, we propose a contrastive quantization-based semantic tokenization approach, named CoST, which harnesses both item relationships and semantic information to learn semantic tokens. Our experimental results highlight the significant impact of semantic tokenization on generative recommendation performance, with CoST achieving up to a 43% improvement in Recall@5 and 44% improvement in NDCG@5 on the MIND dataset over previous baselines.
In this work, the problem of communicating decisions of a classifier over a noisy channel is considered. With machine learning based models being used in variety of time-sensitive applications, transmission of these decisions in a reliable and timely manner is of significant importance. To this end, we study the scenario where a probability vector (representing the decisions of a classifier) at the transmitter, needs to be transmitted over a noisy channel. Assuming that the distortion between the original probability vector and the reconstructed one at the receiver is measured via f-divergence, we study the trade-off between transmission latency and the distortion. We completely analyze this trade-off using uniform, lattice, and sparse lattice-based quantization techniques to encode the probability vector by first characterizing bit budgets for each technique given a requirement on the allowed source distortion. These bounds are then combined with results from finite-blocklength literature to provide a framework for analyzing the effects of both quantization distortion and distortion due to decoding error probability (i.e., channel effects) on the incurred transmission latency. Our results show that there is an interesting interplay between source distortion (i.e., distortion for the probability vector measured via f-divergence) and the subsequent channel encoding/decoding parameters; and indicate that a joint design of these parameters is crucial to navigate the latency-distortion tradeoff. We study the impact of changing different parameters (e.g. number of classes, SNR, source distortion) on the latency-distortion tradeoff and perform experiments on AWGN and fading channels. Our results indicate that sparse lattice-based quantization is the most effective at minimizing latency across various regimes and for sparse, high-dimensional probability vectors (i.e., high number of classes).
Fine-tuning Large Language Models (LLMs) is a common practice to adapt pre-trained models for specific applications. While methods like LoRA have effectively addressed GPU memory constraints during fine-tuning, their performance often falls short, especially in multi-task scenarios. In contrast, Mixture-of-Expert (MoE) models, such as Mixtral 8x7B, demonstrate remarkable performance in multi-task learning scenarios while maintaining a reduced parameter count. However, the resource requirements of these MoEs remain challenging, particularly for consumer-grade GPUs with less than 24GB memory. To tackle these challenges, we propose MixLoRA, an approach to construct a resource-efficient sparse MoE model based on LoRA. MixLoRA inserts multiple LoRA-based experts within the feed-forward network block of a frozen pre-trained dense model and employs a commonly used top-k router. Unlike other LoRA-based MoE methods, MixLoRA enhances model performance by utilizing independent attention-layer LoRA adapters. Additionally, an auxiliary load balance loss is employed to address the imbalance problem of the router. Our evaluations show that MixLoRA improves about 9% accuracy compared to state-of-the-art PEFT methods in multi-task learning scenarios. We also propose a new high-throughput framework to alleviate the computation and memory bottlenecks during the training and inference of MOE models. This framework reduces GPU memory consumption by 40% and token computation latency by 30% during both training and inference.
The LLaMA family, a collection of foundation language models ranging from 7B to 65B parameters, has become one of the most powerful open-source large language models (LLMs) and the popular LLM backbone of multi-modal large language models (MLLMs), widely used in computer vision and natural language understanding tasks. In particular, LLaMA3 models have recently been released and have achieved impressive performance in various domains with super-large scale pre-training on over 15T tokens of data. Given the wide application of low-bit quantization for LLMs in resource-constrained scenarios, we explore LLaMA3's capabilities when quantized to low bit-width. This exploration can potentially provide new insights and challenges for the low-bit quantization of LLaMA3 and other future LLMs, especially in addressing performance degradation issues that suffer in LLM compression. Specifically, we comprehensively evaluate the 10 existing post-training quantization and LoRA fine-tuning (LoRA-FT) methods of LLaMA3 on 1-8 bits and various datasets to reveal the low-bit quantization performance of LLaMA3. To uncover the capabilities of low-bit quantized MLLM, we assessed the performance of the LLaMA3-based LLaVA-Next-8B model under 2-4 ultra-low bits with post-training quantization methods. Our experimental results indicate that LLaMA3 still suffers from non-negligible degradation in linguistic and visual contexts, particularly under ultra-low bit widths. This highlights the significant performance gap at low bit-width that needs to be addressed in future developments. We expect that this empirical study will prove valuable in advancing future models, driving LLMs and MLLMs to achieve higher accuracy at lower bit to enhance practicality. Our project is released on https://github.com/Macaronlin/LLaMA3-Quantization , and quantized models are released at https://huggingface.co/Efficient-ML .
This paper investigates the challenging problem of learned image compression (LIC) with extreme low bitrates. Previous LIC methods based on transmitting quantized continuous features often yield blurry and noisy reconstruction due to the severe quantization loss. While previous LIC methods based on learned codebooks that discretize visual space usually give poor-fidelity reconstruction due to the insufficient representation power of limited codewords in capturing faithful details. We propose a novel dual-stream framework, HyrbidFlow, which combines the continuous-feature-based and codebook-based streams to achieve both high perceptual quality and high fidelity under extreme low bitrates. The codebook-based stream benefits from the high-quality learned codebook priors to provide high quality and clarity in reconstructed images. The continuous feature stream targets at maintaining fidelity details. To achieve the ultra low bitrate, a masked token-based transformer is further proposed, where we only transmit a masked portion of codeword indices and recover the missing indices through token generation guided by information from the continuous feature stream. We also develop a bridging correction network to merge the two streams in pixel decoding for final image reconstruction, where the continuous stream features rectify biases of the codebook-based pixel decoder to impose reconstructed fidelity details. Experimental results demonstrate superior performance across several datasets under extremely low bitrates, compared with existing single-stream codebook-based or continuous-feature-based LIC methods.
In federated learning, particularly in cross-device scenarios, secure aggregation has recently gained popularity as it effectively defends against inference attacks by malicious aggregators. However, secure aggregation often requires additional communication overhead and can impede the convergence rate of the global model, which is particularly challenging in wireless network environments with extremely limited bandwidth. Therefore, achieving efficient communication compression under the premise of secure aggregation presents a highly challenging and valuable problem. In this work, we propose a novel uplink communication compression method for federated learning, named FedMPQ, which is based on multi shared codebook product quantization.Specifically, we utilize updates from the previous round to generate sufficiently robust codebooks. Secure aggregation is then achieved through trusted execution environments (TEE) or a trusted third party (TTP).In contrast to previous works, our approach exhibits greater robustness in scenarios where data is not independently and identically distributed (non-IID) and there is a lack of sufficient public data. The experiments conducted on the LEAF dataset demonstrate that our proposed method achieves 99% of the baseline's final accuracy, while reducing uplink communications by 90-95%
The sim-to-real gap poses a significant challenge in RL-based multi-agent exploration due to scene quantization and action discretization. Existing platforms suffer from the inefficiency in sampling and the lack of diversity in Multi-Agent Reinforcement Learning (MARL) algorithms across different scenarios, restraining their widespread applications. To fill these gaps, we propose MAexp, a generic platform for multi-agent exploration that integrates a broad range of state-of-the-art MARL algorithms and representative scenarios. Moreover, we employ point clouds to represent our exploration scenarios, leading to high-fidelity environment mapping and a sampling speed approximately 40 times faster than existing platforms. Furthermore, equipped with an attention-based Multi-Agent Target Generator and a Single-Agent Motion Planner, MAexp can work with arbitrary numbers of agents and accommodate various types of robots. Extensive experiments are conducted to establish the first benchmark featuring several high-performance MARL algorithms across typical scenarios for robots with continuous actions, which highlights the distinct strengths of each algorithm in different scenarios.