Low-Shot Learning with Imprinted Weights

By Hang Qi, Matthew A. Brown, David G. Lowe. 2017

549 citations recorded by the source corpus.

Human vision is able to immediately recognize novel visual categories after seeing just one or a few training examples. We describe how to add a similar capability to ConvNet classifiers by directly setting the final layer weights from novel training examples during low-shot learning. We call this process weight imprinting as it directly sets weights for a new category based on an appropriately scaled copy of the embedding layer activations for that training example. The imprinting process provides a valuable complement to training with stochastic gradient descent, as it provides immediate good classification performance and an initialization for any further fine-tuning in the future. We show how this imprinting process is related to proxy-based embeddings. However, it differs in that only a single imprinted weight vector is learned for each novel category, rather than relying on a nearest-neighbor distance to training instances as typically used with embedding methods. Our experiments show that using averaging of imprinted weights provides better generalization than using nearest-neighbor instance embeddings.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Meta & Continual Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Meta & Continual Learning

Citations in the corpus

A Closer Look at Few-shot Classification (2019, methodology): “Second, by replacing the linear classiﬁer with a distance-based classiﬁer as used in Gidaris & Komodakis (2018); Qi et al. (2018), the baseline method is surprisingly competitive to current state-of-art meta-learning algorithms.”
Margin-Based Few-Shot Class-Incremental Learning with Class-Level Overfitting Mitigation (2022, methodology): “It can be roughly divided into metric-based methods [32, 28, 39, 25], meta-learning based methods [12, 23, 27] and augmentation-based methods [36, 14, 1].”
Continual Learning through Retrieval and Imagination (2022, methodology): “Although co-sine normalization has been adopted in other visual problem (Gidaris et al. 2018; Wei et al. 2019; Guo et al. 2020; Qi et al. 2018; Hou et al. 2019), we demonstrate that it can effectively eliminate the imbalance in the magnitude of class embeddings and facilitate continual learning…”
Few-Shot Class-Incremental Learning by Sampling Multi-Phase Tasks (2022, background): “There is a similar setting proposed by [71], [72], [73], namely Generalized Few-Shot Learning (GFSL).”
Subspace Regularizers for Few-Shot Class Incremental Learning (2021, methodology): “This version, which we call single-session FSCIL, has been extensively studied by previous work (Qi et al., 2018; Gidaris & Komodakis, 2018; Ren et al., 2019; Yoon et al., 2020).”
An EM Framework for Online Incremental Learning of Semantic Segmentation (2021, neutral): “It is worth noting that we are the first to adopt cosine normalization in the online class incremental learning problem, even though it has been used in other problems [22, 26, 35].”
New Insights on Reducing Abrupt Representation Change in Online Continual Learning (2021, methodology): “…on the incoming data X in . where sim ( a, b ) = exp( a T b τ (cid:107) a (cid:107)(cid:107) b ) computes the exponential cosine similarity between two vectors, with scaling factor τ (Qi et al., 2018; He et al., 2020). while The stream has not ended Here we denote the incoming data x i ∈ X in .”
Incremental Few-shot Text Classification with Multi-round New Classes: Formulation, Dataset and System (2021, positive): “In Ren et al. (2019 Qi et al. (2018), both Gidaris and Komodakis (2018) and Ren et al. (2019) build connections be-tween the representations of base classes and the new classes.”
Semantic-aware Knowledge Distillation for Few-Shot Class-Incremental Learning (2021, methodology): “Method 1-shot 5-shot accuracy ∆ accuracy ∆ Imprint [23] 41.”
A Two-phase Prototypical Network Model for Incremental Few-shot Relation Classification (2020, background): “Currently, several related works (Qi et al., 2018; Gidaris and Komodakis, 2018; Xiang et al., 2019; Ren et al., 2019) are proposed in computer vision field and they concentrate on image classification task.”
Incremental Few-Shot Learning with Attention Attractor Networks (2018, methodology): “On both benchmarks, our best performing model shows a significant margin over the prior works that predict the prototype representation without using an iterative optimization [33, 26, 9].”
Large-Scale Long-Tailed Recognition in an Open World (2019, methodology): “We adopt the cosine classifier [35, 13] to produce the final classification results.”
A Baseline for Few-Shot Image Classification (2019, background): “is indicates that CIFAR-FS, FC-100 and Mini-ImageNet may be good benchmarks for applications with few classes. The hardness metric in (9) naturally builds upon existing ideas in deep metric learning (Qi et al., 2018). We propose it as a means to evaluate few-shot learning algorithms uniformly across”
Frustratingly Simple Few-Shot Object Detection (2020, methodology): “More recently, several (Chen et al., 2019; Gidaris & Komodakis, 2018; Qi et al., 2018) adopt a cosine similarity based classifier to reduce the intraclass variance on the few-shot classification task, which leads to favorable performance compared to many metalearning based approaches.”
Few-Shot Object Detection via Feature Reweighting (2018, methodology): “Weight imprinting [30] sets weights for a new category using a scaled embedding of labeled examples.”
Boosting Few-Shot Visual Learning With Self-Supervision (2019, methodology): “, training images, of a few-shot task [26, 33, 53, 55, 57, 59]; methods learning to map a test example to a class label by accessing memory modules that store training examples for that task [11, 25, 35, 38, 50]; approaches that learn how to generate the weights of a classifier [12, 16, 43, 44, 13] ”
Generalizing from a Few Examples (2019, methodology): “ses by similar features taken from the novel class and reusing their classifier parameters. Only the threshold for classification is adjusted to avoid confusion with those similar classes. Similar to [89] and [90], a pre-trained CNN is adapted to deal with new class in [40]. But instead of solely us”