Kullback-Leiber divergence has been widely used in Knowledge Distillation (KD) to compress Large Language Models (LLMs). Contrary to prior assertions that reverse Kullback-Leibler (RKL) divergence is mode-seeking and thus preferable over the mean-seeking forward Kullback-Leibler (FKL) divergence, this study empirically and theoretically demonstrates that neither mode-seeking nor mean-seeking properties manifest in KD for LLMs. Instead, RKL and FKL are found to share the same optimization objective and both converge after a sufficient number of epochs. However, due to practical constraints, LLMs are seldom trained for such an extensive number of epochs. Meanwhile, we further find that RKL focuses on the tail part of the distributions, while FKL focuses on the head part at the beginning epochs. Consequently, we propose a simple yet effective Adaptive Kullback-Leiber (AKL) divergence method, which adaptively allocates weights to combine FKL and RKL. Metric-based and GPT-4-based evaluations demonstrate that the proposed AKL outperforms the baselines across various tasks and improves the diversity and quality of generated responses.
The widespread use of high-definition screens in edge devices, such as end-user cameras, smartphones, and televisions, is spurring a significant demand for image enhancement. Existing enhancement models often optimize for high performance while falling short of reducing hardware inference time and power consumption, especially on edge devices with constrained computing and storage resources. To this end, we propose Image Color Enhancement Lookup Table (ICELUT) that adopts LUTs for extremely efficient edge inference, without any convolutional neural network (CNN). During training, we leverage pointwise (1x1) convolution to extract color information, alongside a split fully connected layer to incorporate global information. Both components are then seamlessly converted into LUTs for hardware-agnostic deployment. ICELUT achieves near-state-of-the-art performance and remarkably low power consumption. We observe that the pointwise network structure exhibits robust scalability, upkeeping the performance even with a heavily downsampled 32x32 input image. These enable ICELUT, the first-ever purely LUT-based image enhancer, to reach an unprecedented speed of 0.4ms on GPU and 7ms on CPU, at least one order faster than any CNN solution. Codes are available at https://github.com/Stephen0808/ICELUT.
Large Language Models (LLMs) have greatly advanced the natural language processing paradigm. However, the high computational load and huge model sizes pose a grand challenge for deployment on edge devices. To this end, we propose APTQ (Attention-aware Post-Training Mixed-Precision Quantization) for LLMs, which considers not only the second-order information of each layer's weights, but also, for the first time, the nonlinear effect of attention outputs on the entire model. We leverage the Hessian trace as a sensitivity metric for mixed-precision quantization, ensuring an informed precision reduction that retains model performance. Experiments show APTQ surpasses previous quantization methods, achieving an average of 4 bit width a 5.22 perplexity nearly equivalent to full precision in the C4 dataset. In addition, APTQ attains state-of-the-art zero-shot accuracy of 68.24\% and 70.48\% at an average bitwidth of 3.8 in LLaMa-7B and LLaMa-13B, respectively, demonstrating its effectiveness to produce high-quality quantized LLMs.
Various parameter-efficient fine-tuning (PEFT) techniques have been proposed to enable computationally efficient fine-tuning while maintaining model performance. However, existing PEFT methods are still limited by the growing number of trainable parameters with the rapid deployment of Large Language Models (LLMs). To address this challenge, we present LoRETTA, an ultra-parameter-efficient framework that significantly reduces trainable parameters through tensor-train decomposition. Specifically, we propose two methods, named {LoRETTA}$_{adp}$ and {LoRETTA}$_{rep}$. The former employs tensorized adapters, offering a high-performance yet lightweight approach for the fine-tuning of LLMs. The latter emphasizes fine-tuning via weight parameterization with a set of small tensor factors. LoRETTA achieves comparable or better performance than most widely used PEFT methods with up to $100\times$ fewer parameters on the LLaMA-2-7B models. Furthermore, empirical results demonstrate that the proposed method effectively improves training efficiency, enjoys better multi-task learning performance, and enhances the anti-overfitting capability. Plug-and-play codes built upon the Huggingface framework and PEFT library will be released.
Existing approaches to Implicit Neural Representation (INR) can be interpreted as a global scene representation via a linear combination of Fourier bases of different frequencies. However, such universal basis functions can limit the representation capability in local regions where a specific component is unnecessary, resulting in unpleasant artifacts. To this end, we introduce a learnable spatial mask that effectively dispatches distinct Fourier bases into respective regions. This translates into collaging Fourier patches, thus enabling an accurate representation of complex signals. Comprehensive experiments demonstrate the superior reconstruction quality of the proposed approach over existing baselines across various INR tasks, including image fitting, video representation, and 3D shape representation. Our method outperforms all other baselines, improving the image fitting PSNR by over 3dB and 3D reconstruction to 98.81 IoU and 0.0011 Chamfer Distance.
Despite the decomposition of convolutional kernels for lightweight CNNs being well studied, existing works that rely on tensor network diagrams or hyperdimensional abstraction lack geometry intuition. This work devises a new perspective by linking a 3D-reshaped kernel tensor to its various slice-wise and rank-1 decompositions, permitting a straightforward connection between various tensor approximations and efficient CNN modules. Specifically, it is discovered that a pointwise-depthwise-pointwise (PDP) configuration constitutes a viable construct for lightweight CNNs. Moreover, a novel link to the latest ShiftNet is established, inspiring a first-ever shift layer pruning that achieves nearly 50% compression with < 1% drop in accuracy for ShiftResNet.
Conventional super-resolution (SR) schemes make heavy use of convolutional neural networks (CNNs), which involve intensive multiply-accumulate (MAC) operations, and require specialized hardware such as graphics processing units. This contradicts the regime of edge AI that often runs on devices strained by power, computing, and storage resources. Such a challenge has motivated a series of lookup table (LUT)-based SR schemes that employ simple LUT readout and largely elude CNN computation. Nonetheless, the multi-megabyte LUTs in existing methods still prohibit on-chip storage and necessitate off-chip memory transport. This work tackles this storage hurdle and innovates hundred-kilobyte LUT (HKLUT) models amenable to on-chip cache. Utilizing an asymmetric two-branch multistage network coupled with a suite of specialized kernel patterns, HKLUT demonstrates an uncompromising performance and superior hardware efficiency over existing LUT schemes.
Most deep neural networks (DNNs) consist fundamentally of convolutional and/or fully connected layers, wherein the linear transform can be cast as the product between a filter matrix and a data matrix obtained by arranging feature tensors into columns. The lately proposed deformable butterfly (DeBut) decomposes the filter matrix into generalized, butterflylike factors, thus achieving network compression orthogonal to the traditional ways of pruning or low-rank decomposition. This work reveals an intimate link between DeBut and a systematic hierarchy of depthwise and pointwise convolutions, which explains the empirically good performance of DeBut layers. By developing an automated DeBut chain generator, we show for the first time the viability of homogenizing a DNN into all DeBut layers, thus achieving an extreme sparsity and compression. Various examples and hardware benchmarks verify the advantages of All-DeBut networks. In particular, we show it is possible to compress a PointNet to < 5% parameters with < 5% accuracy drop, a record not achievable by other compression schemes.
Deep neural networks (DNNs) are incredibly vulnerable to crafted, imperceptible adversarial perturbations. While adversarial training (AT) has proven to be an effective defense approach, the AT mechanism for robustness improvement is not fully understood. This work investigates AT from a spectral perspective, adding new insights to the design of effective defenses. In particular, we show that AT induces the deep model to focus more on the low-frequency region, which retains the shape-biased representations, to gain robustness. Further, we find that the spectrum of a white-box attack is primarily distributed in regions the model focuses on, and the perturbation attacks the spectral bands where the model is vulnerable. Based on this observation, to train a model tolerant to frequency-varying perturbation, we propose a spectral alignment regularization (SAR) such that the spectral output inferred by an attacked adversarial input stays as close as possible to its natural input counterpart. Experiments demonstrate that SAR and its weight averaging (WA) extension could significantly improve the robust accuracy by 1.14% ~ 3.87% relative to the standard AT, across multiple datasets (CIFAR-10, CIFAR-100 and Tiny ImageNet), and various attacks (PGD, C&W and Autoattack), without any extra data.
The beam squint effect, which manifests in different steering matrices in different sub-bands, has been widely considered a challenge in millimeter wave (mmWave) multiinput multi-output (MIMO) channel estimation. Existing methods either require specific forms of the precoding/combining matrix, which restrict their general practicality, or simply ignore the beam squint effect by only making use of a single sub-band for channel estimation. Recognizing that different steering matrices are coupled by the same set of unknown channel parameters, this paper proposes to exploit the common sparsity structure of the virtual channel model so that signals from different subbands can be jointly utilized to enhance the performance of channel estimation. A probabilistic model is built to induce the common sparsity in the spatial domain, and the first-order Taylor expansion is adopted to get rid of the grid mismatch in the dictionaries. To learn the model parameters, a variational expectation-maximization (EM) algorithm is derived, which automatically obtains the balance between the likelihood function and the common sparsity prior information, and is applicable to arbitrary forms of precoding/combining matrices. Simulation results show the superior estimation accuracy of the proposed algorithm over existing methods under different noise powers and system configurations.