CNN LEGO: Disassembling and Assembling Convolutional Neural Network

Convolutional Neural Network (CNN), which mimics human visual perception mechanism, has been successfully used in many computer vision areas. Some psychophysical studies show that the visual perception mechanism synchronously processes the form, color, movement, depth, etc., in the initial stage [7,20] and then integrates all information for final recognition [38]. What's more, the human visual system [20] contains different subdivisions or different tasks. Inspired by the above visual perception mechanism, we investigate a new task, termed as Model Disassembling and Assembling (MDA-Task), which can disassemble the deep models into independent parts and assemble those parts into a new deep model without performance cost like playing LEGO toys. To this end, we propose a feature route attribution technique (FRAT) for disassembling CNN classifiers in this paper. In FRAT, the positive derivatives of predicted class probability w.r.t. the feature maps are adopted to locate the critical features in each layer. Then, relevance analysis between the critical features and preceding/subsequent parameter layers is adopted to bridge the route between two adjacent parameter layers. In the assembling phase, class-wise components of each layer are assembled into a new deep model for a specific task. Extensive experiments demonstrate that the assembled CNN classifier can achieve close accuracy with the original classifier without any fine-tune, and excess original performance with one-epoch fine-tune. What's more, we also conduct massive experiments to verify the broad application of MDA-Task on model decision route visualization, model compression, knowledge distillation, transfer learning, incremental learning, and so on.

Root-aligned SMILES for Molecular Retrosynthesis Prediction

Retrosynthesis prediction is a fundamental problem in organic synthesis, where the task is to discover precursor molecules that can be used to synthesize a target molecule. A popular paradigm of existing computational retrosynthesis methods formulate retrosynthesis prediction as a sequence-to-sequence translation problem, where the typical SMILES representations are adopted for both reactants and products. However, the general-purpose SMILES neglects the characteristics of retrosynthesis that 1) the search space of the reactants is quite huge, and 2) the molecular graph topology is largely unaltered from products to reactants, resulting in the suboptimal performance of SMILES if straightforwardly applied. In this article, we propose the root-aligned SMILES~(R-SMILES), which specifies a tightly aligned one-to-one mapping between the product and the reactant SMILES, to narrow the string representation discrepancy for more efficient retrosynthesis. As the minimum edit distance between the input and the output is significantly decreased with the proposed R-SMILES, the computational model is largely relieved from learning the complex syntax and dedicated to learning the chemical knowledge for retrosynthesis. We compare the proposed R-SMILES with various state-of-the-art baselines on different benchmarks and show that it significantly outperforms them all, demonstrating the superiority of the proposed method.

Meta-attention for ViT-backed Continual Learning

Continual learning is a longstanding research topic due to its crucial role in tackling continually arriving tasks. Up to now, the study of continual learning in computer vision is mainly restricted to convolutional neural networks (CNNs). However, recently there is a tendency that the newly emerging vision transformers (ViTs) are gradually dominating the field of computer vision, which leaves CNN-based continual learning lagging behind as they can suffer from severe performance degradation if straightforwardly applied to ViTs. In this paper, we study ViT-backed continual learning to strive for higher performance riding on recent advances of ViTs. Inspired by mask-based continual learning methods in CNNs, where a mask is learned per task to adapt the pre-trained ViT to the new task, we propose MEta-ATtention (MEAT), i.e., attention to self-attention, to adapt a pre-trained ViT to new tasks without sacrificing performance on already learned tasks. Unlike prior mask-based methods like Piggyback, where all parameters are associated with corresponding masks, MEAT leverages the characteristics of ViTs and only masks a portion of its parameters. It renders MEAT more efficient and effective with less overhead and higher accuracy. Extensive experiments demonstrate that MEAT exhibits significant superiority to its state-of-the-art CNN counterparts, with 4.0~6.0% absolute boosts in accuracy. Our code has been released at https://github.com/zju-vipa/MEAT-TIL.

Knowledge Amalgamation for Object Detection with Transformers

Knowledge amalgamation (KA) is a novel deep model reusing task aiming to transfer knowledge from several well-trained teachers to a multi-talented and compact student. Currently, most of these approaches are tailored for convolutional neural networks (CNNs). However, there is a tendency that transformers, with a completely different architecture, are starting to challenge the domination of CNNs in many computer vision tasks. Nevertheless, directly applying the previous KA methods to transformers leads to severe performance degradation. In this work, we explore a more effective KA scheme for transformer-based object detection models. Specifically, considering the architecture characteristics of transformers, we propose to dissolve the KA into two aspects: sequence-level amalgamation (SA) and task-level amalgamation (TA). In particular, a hint is generated within the sequence-level amalgamation by concatenating teacher sequences instead of redundantly aggregating them to a fixed-size one as previous KA works. Besides, the student learns heterogeneous detection tasks through soft targets with efficiency in the task-level amalgamation. Extensive experiments on PASCAL VOC and COCO have unfolded that the sequence-level amalgamation significantly boosts the performance of students, while the previous methods impair the students. Moreover, the transformer-based students excel in learning amalgamated knowledge, as they have mastered heterogeneous detection tasks rapidly and achieved superior or at least comparable performance to those of the teachers in their specializations.

Imbalanced Sample Generation and Evaluation for Power System Transient Stability Using CTGAN

Although deep learning has achieved impressive advances in transient stability assessment of power systems, the insufficient and imbalanced samples still trap the training effect of the data-driven methods. This paper proposes a controllable sample generation framework based on Conditional Tabular Generative Adversarial Network (CTGAN) to generate specified transient stability samples. To fit the complex feature distribution of the transient stability samples, the proposed framework firstly models the samples as tabular data and uses Gaussian mixture models to normalize the tabular data. Then we transform multiple conditions into a single conditional vector to enable multi-conditional generation. Furthermore, this paper introduces three evaluation metrics to verify the quality of generated samples based on the proposed framework. Experimental results on the IEEE 39-bus system show that the proposed framework effectively balances the transient stability samples and significantly improves the performance of transient stability assessment models.

Up to 100x Faster Data-free Knowledge Distillation

Data-free knowledge distillation (DFKD) has recently been attracting increasing attention from research communities, attributed to its capability to compress a model only using synthetic data. Despite the encouraging results achieved, state-of-the-art DFKD methods still suffer from the inefficiency of data synthesis, making the data-free training process extremely time-consuming and thus inapplicable for large-scale tasks. In this work, we introduce an efficacious scheme, termed as FastDFKD, that allows us to accelerate DFKD by a factor of orders of magnitude. At the heart of our approach is a novel strategy to reuse the shared common features in training data so as to synthesize different data instances. Unlike prior methods that optimize a set of data independently, we propose to learn a meta-synthesizer that seeks common features as the initialization for the fast data synthesis. As a result, FastDFKD achieves data synthesis within only a few steps, significantly enhancing the efficiency of data-free training. Experiments over CIFAR, NYUv2, and ImageNet demonstrate that the proposed FastDFKD achieves 10$\times$ and even 100$\times$ acceleration while preserving performances on par with state of the art.

A Survey of Deep Learning for Low-Shot Object Detection

Object detection is a fundamental task in computer vision and image processing. Current deep learning based object detectors have been highly successful with abundant labeled data. But in real life, it is not guaranteed that each object category has enough labeled samples for training. These large object detectors are easy to overfit when the training data is limited. Therefore, it is necessary to introduce few-shot learning and zero-shot learning into object detection, which can be named low-shot object detection together. Low-Shot Object Detection (LSOD) aims to detect objects from a few or even zero labeled data, which can be categorized into few-shot object detection (FSOD) and zero-shot object detection (ZSD), respectively. This paper conducts a comprehensive survey for deep learning based FSOD and ZSD. First, this survey classifies methods for FSOD and ZSD into different categories and discusses the pros and cons of them. Second, this survey reviews dataset settings and evaluation metrics for FSOD and ZSD, then analyzes the performance of different methods on these benchmarks. Finally, this survey discusses future challenges and promising directions for FSOD and ZSD.

Bootstrapping ViTs: Towards Liberating Vision Transformers from Pre-training

Recently, vision Transformers (ViTs) are developing rapidly and starting to challenge the domination of convolutional neural networks (CNNs) in the realm of computer vision (CV). With the general-purpose Transformer architecture for replacing the hard-coded inductive biases of convolution, ViTs have surpassed CNNs, especially in data-sufficient circumstances. However, ViTs are prone to over-fit on small datasets and thus rely on large-scale pre-training, which expends enormous time. In this paper, we strive to liberate ViTs from pre-training by introducing CNNs' inductive biases back to ViTs while preserving their network architectures for higher upper bound and setting up more suitable optimization objectives. To begin with, an agent CNN is designed based on the given ViT with inductive biases. Then a bootstrapping training algorithm is proposed to jointly optimize the agent and ViT with weight sharing, during which the ViT learns inductive biases from the intermediate features of the agent. Extensive experiments on CIFAR-10/100 and ImageNet-1k with limited training data have shown encouraging results that the inductive biases help ViTs converge significantly faster and outperform conventional CNNs with even fewer parameters.

Model Doctor: A Simple Gradient Aggregation Strategy for Diagnosing and Treating CNN Classifiers

Recently, Convolutional Neural Network (CNN) has achieved excellent performance in the classification task. It is widely known that CNN is deemed as a 'black-box', which is hard for understanding the prediction mechanism and debugging the wrong prediction. Some model debugging and explanation works are developed for solving the above drawbacks. However, those methods focus on explanation and diagnosing possible causes for model prediction, based on which the researchers handle the following optimization of models manually. In this paper, we propose the first completely automatic model diagnosing and treating tool, termed as Model Doctor. Based on two discoveries that 1) each category is only correlated with sparse and specific convolution kernels, and 2) adversarial samples are isolated while normal samples are successive in the feature space, a simple aggregate gradient constraint is devised for effectively diagnosing and optimizing CNN classifiers. The aggregate gradient strategy is a versatile module for mainstream CNN classifiers. Extensive experiments demonstrate that the proposed Model Doctor applies to all existing CNN classifiers, and improves the accuracy of $16$ mainstream CNN classifiers by 1%-5%.