United States

Transformer and its variants achieve excellent results in various computer vision and natural language processing tasks,  but high computational costs and reliance on large training datasets restrict their deployment in resource-constrained settings. Low-rank approximation of model weights has been effective in compressing CNN models, but its application to transformers has been less explored and is less effective. Existing methods require the complete dataset to fine-tune compressed models, which are both time-consuming and data-hungry. This paper reveals that the features (i.e., activations) are low-rank, but model weights are surprisingly not low-rank. Hence, AAFM is proposed, which adaptively determines the compressed model structure and locally compresses each linear layer&#39;s output features rather than the model weights. A second stage, GFM, optimizes the entire compressed network holistically. Both AAFM and GFM only use few training samples without labels, that is, they are few-shot, unsupervised, fast and effective. For example, with only 2K images without labels, 33% of the parameters are removed in DeiT-B with 18.8% relative throughput increase, but only a 0.23% accuracy loss for ImageNet recognition. The proposed methods are successfully applied to the language modeling task in NLP, too. Besides, the few-shot compressed models generalize well in downstream tasks.

AAAI 2023

Compressing Transformers: Features Are Low-Rank, but Weights Are Not!

learning on the edge & model compression

deep neural network algorithms

representation learning for vision

language models

The purpose of the AAAI conference series is to promote research in Artificial Intelligence (AI) and foster scientific exchange between researchers, practitioners, scientists, students, and engineers across the entirety of AI and its affiliated disciplines. AAAI-23 is the Thirty-Seventh AAAI Conference on Artificial Intelligence. The theme of this conference is to create collaborative bridges within and beyond AI. Like previous AAAI conferences, AAAI-23 will feature technical paper presentations, special tracks, invited speakers, workshops, tutorials, poster sessions, senior member presentations, competitions, and exhibit programs, and two new activities: a Bridge Program and a Lab Program. Many of these activities are tailored to the theme of bridges and all are selected according to the highest standards, with additional programs for students and young researchers. 
AAAI is providing you with a conference planner, which you can use to help organize your itinerary of activities. This includes talks to attend in person, talks to attend remotely, breaks with colleagues and your site seeing activities. To access this conference planner, please go to [https://aaai-2023.takemobi.io](https://aaai-2023.takemobi.io).

In order to access this site, you need to register. If you haven't already, please register [here](https://aaai.org/Conferences/AAAI-23/registration/).


The purpose of the AAAI conference series is to promote research in Artificial Intelligence (AI) and foster scientific exchange between researchers, practitioners, scientists, students, and engineers across the entirety of AI and its affiliated disciplines.

Transformer and its variants achieve excellent results in various computer vision and natural language processing tasks,  but high computational costs and reliance on large training datasets restrict their deployment in resource-constrained settings. Low-rank approximation of model weights has been effective in compressing CNN models, but its application to transformers has been less explored and is less effective. Existing methods require the complete dataset to fine-tune compressed models, which are both time-consuming and data-hungry. This paper reveals that the features (i.e., activations) are low-rank, but model weights are surprisingly not low-rank. Hence, AAFM is proposed, which adaptively determines the compressed model structure and locally compresses each linear layer's output features rather than the model weights. A second stage, GFM, optimizes the entire compressed network holistically. Both AAFM and GFM only use few training samples without labels, that is, they are few-shot, unsupervised, fast and effective. For example, with only 2K images without labels, 33% of the parameters are removed in DeiT-B with 18.8% relative throughput increase, but only a 0.23% accuracy loss for ImageNet recognition. The proposed methods are successfully applied to the language modeling task in NLP, too. Besides, the few-shot compressed models generalize well in downstream tasks.

poster

Recent deep learning methods have achieved promising results in image shadow removal. However, most of the existing approaches focus on working locally within shadow and non-shadow regions, resulting in severe artifacts around the shadow boundaries as well as inconsistent illumination between shadow and non-shadow regions. It is still challenging for the deep shadow removal model to exploit the global contextual correlation between shadow and non-shadow regions. In this work, we first propose a Retinex-based shadow model, from which we derive a novel transformer-based network, dubbed \textit{ShandowFormer}, to exploit non-shadow regions to help shadow region restoration. A multi-scale channel attention framework is employed to hierarchically capture the global information. Based on that, we propose a Shadow-Interaction Module (SIM) with Shadow-Interaction Attention (SIA) in the bottleneck stage to effectively model the context correlation between shadow and non-shadow regions. We conduct extensive experiments on three popular public datasets, including ISTD, ISTD+, and SRD, 
to evaluate the proposed method. Our method achieves state-of-the-art performance by using up to $\mathbf{150\times}$ fewer model parameters.



ShadowFormer: Global Context Helps Shadow Removal

Low-light images suffer severe degradation of low lightness and noise corruption, causing unsatisfactory visual quality and visual recognition performance. To solve this problem while meeting the unavailability of paired datasets in widerange scenarios, unsupervised low-light image enhancement (ULLIE) techniques have been developed. However, these  methods are primarily guided to alleviate the degradation effect on visual quality rather than semantic levels, hence limiting their performance in visual recognition tasks. To this end, we propose to learn a Semantic Degradation-Aware Guidance (SDAG) that perceives the low-light degradation effect on semantic levels in a self-supervised manner, which is further
utilized to guide the ULLIE methods. The proposed SDAG utilizes the low-light degradation factors as augmented signals to degrade the low-light images, and then capture their degradation effect on semantic levels. Specifically, our SDAG
employs the subsequent pre-trained recognition model extractor to extract semantic representations, and then learns to self-reconstruct the enhanced low-light image and its augmented degraded images. By constraining the relative reconstruction effect between the original enhanced image and the augmented formats, our SDAG learns to be aware of the degradation effect on semantic levels in a relative comparison manner. Moreover, our SDAG is general and can be plugged into the training paradigm of the existing ULLIE methods. Extensive experiments demonstrate its effectiveness for improving the ULLIE approaches on the downstream recognition tasks while maintaining a competitive visual quality.

Learning Semantic Degradation-Aware Guidance for Recognition-Driven Unsupervised Low-Light Image Enhancement

Restricted by the ability of depth perception, all Multi-view 3D Object Detection methods fall into the bottleneck of depth accuracy. By constructing temporal stereo, depth estimation is quite reliable in indoor scenarios. However, there are two difficulties in directly integrating temporal stereo into outside Multi-view 3D Object Detectors: 1) The construction of temporal stereos for all perspectives resulted in high computing costs. 2) difficult outdoor situations. In this study, we propose an effective method for creating temporal stereo by dynamically determining the center and range of the temporal stereo. The most confident center is found using the EM algorithm. our method also has the ability to deal with scenarios that indoor temporal stereo methods are unable to handle. In addition, a more reasonable NMS method is proposed for better filtering proposes.

BEVStereo: Enhancing Depth Estimation in Multi-view 3D Object Detection with Temporal Stereo

As one of the basic while vital technologies for HD map construction, 3D lane detection is still an open problem due to varying visual conditions, complex typologies, and strict demands for precision. In this paper, an end-to-end flexible and hierarchical lane detector is proposed to precisely predict 3D lane lines from point clouds. Specifically, we design a hierarchical network predicting flexible representations of lane shapes at different levels, simultaneously collecting global instance semantics and avoiding local errors. In the global scope, we propose to regress parametric curves w.r.t adaptive axes that help to make more robust predictions towards complex scenes, while in the local vision the structure of lane segment is detected in each of the dynamic anchor cells sampled along the global predicted curves. Moreover, corresponding global and local shape matching losses and anchor cell generation strategies are designed. Experiments on two datasets show that we overwhelm current top methods under high precision standards, and full ablation studies also verify each part of our method. Our codes will be released at https://github.com/Doo-do/FHLD.

Flexible 3D Lane Detection by Hierarchical Shape Matching

Online learning with feature spaces that are not fixed but can
vary over time renders a seemingly flexible learning paradigm
thus has drawn much attention. Unfortunately, two restrictions
prohibit a ubiquitous application of this learning paradigm in
practice. First, whereas prior studies mainly assume a homogenous feature type, data streams generated from real applications can be heterogeneous in which Boolean, ordinal, and
continuous co-exist. Existing methods that prescribe parametric distributions such as Gaussians would not suffice to model
the correlation among such mix-typed features. Second, while
full supervision seems to be a default setup, providing labels
to all arriving data instances over a long time span is tangibly
onerous, laborious, and economically unsustainable. Alas, a
semi-supervised online learner that can deal with mix-typed,
varying feature spaces is still missing. To fill the gap, this paper explores a novel problem, named Online Semi-supervised
Learning with Mix-typed streaming Features (OSLMF), which
strives to relax the restrictions on the feature type and supervision information. Our key idea to solve the new problem
is to leverage copula model to align the data instances with
different feature spaces so to make their distance measurable.
A geometric structure underlying data instances is then established in an online fashion based on their distances, through
which the limited labeling information is propagated, from the
scarce labeled instances to their close neighbors. Theoretical
analysis and experimental results are documented to evidence
the viability and effectiveness of our proposed approach.


Online Semi-Supervised Learning with Mix-Typed Streaming Features

Imitation learning (IL) has recently shown impressive performance in training a reinforcement learning agent with human demonstrations, eliminating the difficulty of designing elaborate reward functions in complex environments. However, most IL methods work under the assumption of the optimality of the demonstrations and thus cannot learn policies to surpass the demonstrators. Some methods have been investigated to obtain better-than-demonstration (BD) performance with inner human feedback or preference labels. In this paper, we propose a method to learn rewards from suboptimal demonstrations via a weighted preference learning technique (LERP). Specifically, we first formulate the suboptimality of demonstrations as the inaccurate estimation of rewards. The inaccuracy is modeled with a reward noise random variable following the Gumbel distribution. Moreover, we derive an upper bound of the expected return with different noise coefficients and propose a theorem to surpass the demonstrations. Unlike existing literature, our  analysis does not depend on the linear reward constraint. Consequently, we develop a BD model with a weighted preference learning technique. Experimental results on continuous control and high-dimensional discrete control tasks show the superiority of our LERP method over other state-of-the-art BD  methods.

Learning Noise-Induced Reward Functions for Surpassing Demonstrations in Imitation Learning

Recent years, graph contrastive learning (GCL), which aims to learn representations from unlabeled graphs, has made great progress. However, the existing GCL methods mostly adopt human-designed graph augmentations, which are sensitive to various graph datasets. In addition, the contrastive losses originally developed in computer vision have been directly applied to graph data, where the neighboring nodes are regarded as negatives and consequently pushed far apart from the anchor. However, this is contradictory with the homophily assumption of net-works that connected nodes often belong to the same class and should be close to each other. In this work, we propose an end-to-end automatic GCL method, named NCLA to apply neighbor contrastive learning on learnable graph augmentation. Several graph augmented views with adaptive topology are automatically learned by the multi-head graph attention mechanism, which can be compatible with various graph datasets without prior domain knowledge. In addition, a neighbor contrastive loss is devised to allow multiple positives per anchor by taking network topology as the supervised signals. Both augmentations and embeddings are learned end-to-end in the pro-posed NCLA. Extensive experiments on the benchmark datasets demonstrate that NCLA yields the state-of-the-art node classification performance on self-supervised GCL and even exceeds the supervised ones, when the labels are extremely limited. Our code is released at https://github.com/shenxiaocam/NCLA.

Neighbor Contrastive Learning on Learnable Graph Augmentation

Designing a point cloud upsampler, which aims to generate a clean and dense point cloud given a sparse point representation, is a fundamental and challenging problem in computer vision. A line of attempts achieves this goal by establishing a point-to-point mapping function via deep neural networks. However, these approaches are prone to produce outlier points due to the lack of explicit surface-level constraints. To solve this problem, we introduce a novel surface regularizer into the upsampler network by forcing the neural network to learn the underlying parametric surface represented by bicubic functions and rotation functions, where the new generated points are then constrained on the underlying surface. These designs are integrated into two different networks for two tasks that take advantages of upsampling layers -- point cloud upsampling and point cloud completion for evaluation. The state-of-the-art experimental results on both tasks demonstrate the effectiveness of the proposed method. The implementation code will be available at https://github.com/corecai163/PSCU.

Parametric Surface Constrained Upsampler Network for Point Cloud

Capsule neural networks replace simple, scalar-valued neurons with vector-valued capsules. They are motivated by the pattern recognition system in the human brain, where complex objects are decomposed into a hierarchy of simpler object parts. Such a hierarchy is referred to as a parse-tree. Conceptually, capsule neural networks have been defined to mimic this behavior. The capsule neural network (CapsNet), by Sabour, Frosst, and Hinton, is the first actual implementation of the conceptual idea of capsule neural networks. CapsNets achieved state-of-the-art performance on simple image recognition tasks with fewer parameters and greater robustness to affine transformations than comparable approaches. This sparked extensive follow-up research. However, despite major efforts, no work was able to scale the CapsNet architecture to more reasonable-sized datasets. Here, we provide a reason for this failure and argue that it is most likely not possible to scale CapsNets beyond toy examples. In particular, we show that the concept of a parse-tree, the main idea behind capsule neuronal networks, is not present in CapsNets. We also show theoretically and experimentally that CapsNets suffer from a vanishing gradient problem that results in the starvation of many capsules during training.

Why Capsule Neural Networks Do Not Scale: Challenging the Dynamic Parse-Tree Assumption

Image guidance is an effective strategy for depth super-resolution. Generally, most existing methods employ hand-crafted operators to decompose the high-frequency (HF) and low-frequency (LF) ingredients from low-resolution depth maps and guide the HF ingredients by directly concatenating them with image features. However, the hand-designed operators usually cause inferior HF maps (e.g., distorted or structurally missing) due to the diverse appearance of complex depth maps. Moreover, the direct concatenation often results in weak guidance because not all image features have a positive effect on the HF maps. In this paper, we develop a recurrent structure attention guided (RSAG) framework, consisting of two important parts. First, we introduce a deep contrastive network with multi-scale filters for adaptive frequency-domain separation, which adopts contrastive networks from large filters to small ones to calculate the pixel contrasts for adaptive high-quality HF predictions. Second, instead of the coarse concatenation guidance, we propose a recurrent structure attention block, which iteratively utilizes the latest depth estimation and the image features to jointly select clear patterns and boundaries, aiming at providing refined guidance for accurate depth recovery. In addition, we fuse the features of HF maps to enhance the edge structures in the decomposed LF maps. Extensive experiments show that our approach obtains superior performance compared with state-of-the-art depth super-resolution methods.

Compressing Transformers: Features Are Low-Rank, but Weights Are Not!

Downloads

Next from AAAI 2023

ShadowFormer: Global Context Helps Shadow Removal

Similar lecture

EffConv: Efficient Learning of Kernel Sizes for Convolution Layers of CNNs