Few-Shot Learning Through an Information Retrieval Lens

By Eleni Triantafillou, Richard S. Zemel, Raquel Urtasun. 2017

252 citations recorded by the source corpus.

Few-shot learning refers to understanding new concepts from only a few examples. We propose an information retrieval-inspired approach for this problem that is motivated by the increased importance of maximally leveraging all the available information in this low-data regime. We define a training objective that aims to extract as much information as possible from each training batch by effectively optimizing over all relative orderings of the batch points simultaneously. In particular, we view each batch point as a `query' that ranks the remaining ones based on its predicted relevance to them and we define a model within the framework of structured prediction to optimize mean Average Precision over these rankings. Our method achieves impressive results on the standard few-shot classification benchmarks while is also capable of few-shot retrieval.

Subfields: Artificial Intelligence, Natural Language Processing, Representation & Self-Supervised 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 Natural Language Processing

In Representation & Self-Supervised Learning

Citations in the corpus

Rethinking Few-Shot Image Classification: a Good Embedding Is All You Need? (2020, background): “Recently, significant advances [56, 55, 53, 12, 46, 48, 57, 34, 44, 62, 26, 28] have been made to tackle this problem using the ideas of meta-learning or “learning to learn”.”
Learning Deep Disentangled Embeddings with the F-Statistic Loss (2018, methodology): “…focusing on deep embeddings which are evaluated via few-shot learning (e.g., Ustinova & Lempitsky, 2016; Schroff et al., 2015; Yi et al., 2014b) and researchers focusing on few-shot learning who have found deep embeddings to be a useful method (e.g., Snell et al., 2017; Triantaﬁllou et al., 2017).”
Recasting Gradient-Based Meta-Learning as Hierarchical Bayes (2018, methodology): “Modern methods accomplish few-shot learning either through the design of network architectures that ingest the few-shot training samples directly (e.g., Koch, 2015; Vinyals et al., 2016; Snell et al., 2017; Hariharan & Girshick, 2017; Triantafillou et al., 2017), or formulating the problem as one of”
Continual Training of Language Models for Few-Shot Learning (2022, methodology): “…methods to overcome over-ﬁtting fall in three main families: (i) model-based methods try to reduce the hypothesis space of the few-shot task (Triantaﬁllou et al., 2017; Hu et al., 2018), (ii) data-based methods try to augment additional data to the few-shot set (Benaim and Wolf, 2018; Gao…”
Continual Few-shot Relation Learning via Embedding Space Regularization and Data Augmentation (2022, methodology): “…and Wolf, 2018; Gao et al., 2020); 160 (ii) model-based methods reduce the hypothesis 161 space using prior knowledge (Rezende et al., 2016; 162 Triantafillou et al., 2017; Hu et al., 2018); and 163 (iii) algorithm-based methods try to find a more 164 suitable strategy to search for the best…”
LFPT5: A Unified Framework for Lifelong Few-shot Language Learning Based on Prompt Tuning of T5 (2021, methodology): “Existing methods to overcome over-2 ﬁtting include: ( i ) model-based methods that explore how to reduce the hypothesis space of the few-shot task (Triantaﬁllou et al., 2017; Hu et al., 2018), ( ii ) data-based methods that try to augment additional data to the few-shot set (Benaim & Wolf, 2018;…”
Pre-training Tasks for Embedding-based Large-scale Retrieval (2020, background): “In the literature, it is also known as the Siamese network (Das et al., 2016; Triantaﬁllou et al., 2017) or dual-encoder model (Cer et al., 2018; Mazar ´ e et al., 2018).”
Learning Adaptive Classifiers Synthesis for Generalized Few-Shot Learning (2019, background): “FSL has achieved promising results in various domains such as visual recognition [13, 38, 57], domain adaptation [14, 24], neural machine translation [19], data compression [64], and density estimation [48].”
Generalizing from a Few Examples (2019, methodology): “The two most popular applications are character recognition [14, 36, 37, 65, 70, 96, 113, 114, 119, 121, 130, 138, 146] and image classification [37, 70, 96, 106, 119, 121, 127, 130, 132, 138, 142, 143, 149].”
Diverse Few-Shot Text Classification with Multiple Metrics (2018, methodology): “For example, the methods in (Triantafillou et al., 2017) can be viewed as training meta-models from any sampled batches from one single meta-training dataset.”
Task-Embedded Control Networks for Few-Shot Imitation Learning (2018, background): “One-shot and few-shot learning is the paradigm of learning from a small number of examples at test time, and has been widely studied in the image recognition community [17, 18, 19, 20, 21, 22].”
Few-Shot Learning via Embedding Adaptation With Set-to-Set Functions (2018, methodology): “On CUB, we randomly sampled 100 species as SEEN classes, another two 50 species are used as two UNSEEN sets for model validation and evaluation [23].”
Characterizing Generalization under Out-Of-Distribution Shifts in Deep Metric Learning (2021, background): “As a remedy few-shot learning [60, 67, 14, 48, 32, 6, 66] assumes few samples of the test distribution to to be available during training, i.”
Transductive Episodic-Wise Adaptive Metric for Few-Shot Learning (2019, methodology): “And all images are cropped with the provided bounding box [33].”
Adaptive Subspaces for Few-Shot Learning (2020, unknown): “Inspired by such learning abilities, several approaches are developed to study learning from limited samples [3–12].”
Reconciling meta-learning and continual learning with online mixtures of tasks (2018, unknown): “—— prototypical networks(Snell et al., 2017)c 46.61 ± 0.78 MAML(Finn et al., 2017) 48.70 ± 1.84 LLAMA(Grant et al., 2018) 49.40 ± 1.83 KNN + GNN embedding(Garcia and Bruna, 2017) 49.44 ± 0.28 mAP-DLM(Triantaﬁllou et al., 2017) 49.82 ± 0.78 fwCNN (Hebb)(Munkhdalai and Trischler, 2018) 50.21 ± 0.37 GN”
Adapted Deep Embeddings: A Synthesis of Methods for k-Shot Inductive Transfer Learning (2018, methodology): “Embeddings form the basis of some methods [8, 12, 14, 29, 31, 33].”
The Variational Homoencoder: Learning to learn high capacity generative models from few examples (2018, unknown): “…Rescale only: Discriminative models , log q ( y ; x, X, Y ) Matching Networks [21] 93.8% 98.7% Convnet with memory module [9] 95.0% 98.6% mAP-DLM [20] 95.4% 98.6% Model-Agnostic Meta-learning [4] 95.8% 98.9 % Prototypical Networks [19] 96.0 % 98.9 % (VHE, within-alphabet 1 ) 81.3% 90.3% All…”