Learning from One and Only One Shot

TL;DR

The paper addresses data-scarce generalization and interpretability by modeling innate human priors via a distortable canvas to learn a non-Euclidean general-appearance similarity. It introduces dual distance measures $\mathcal{D}_C$ and $\mathcal{D}_V$, optimized with Abstracted Multi-Level Gradient Descent (AMGD) and anchor grids to yield interpretable transformation flows. With no pretraining, a $1$-NN classifier in the learned space achieves state-of-the-art results in tiny-data regimes on MNIST/EMNIST and near-human performance on Omniglot and QuickDraw one-shot tasks, and extends to unsupervised archetype generation. The approach offers a human-like, data-efficient path toward interpretable similarity learning that can complement or replace heavily pretrained models in data-sparse domains and provide insights into visual abstraction and cognition.

Abstract

Humans can generalize from only a few examples and from little pretraining on similar tasks. Yet, machine learning (ML) typically requires large data to learn or pre-learn to transfer. Motivated by nativism and artificial general intelligence, we directly model human-innate priors in abstract visual tasks such as character and doodle recognition. This yields a white-box model that learns general-appearance similarity by mimicking how humans naturally ``distort'' an object at first sight. Using just nearest-neighbor classification on this cognitively-inspired similarity space, we achieve human-level recognition with only $1$--$10$ examples per class and no pretraining. This differs from few-shot learning that uses massive pretraining. In the tiny-data regime of MNIST, EMNIST, Omniglot, and QuickDraw benchmarks, we outperform both modern neural networks and classical ML. For unsupervised learning, by learning the non-Euclidean, general-appearance similarity space in a $k$-means style, we achieve multifarious visual realizations of abstract concepts by generating human-intuitive archetypes as cluster centroids.

Figures (8)

• Figure 1: Canvas transformations (a), transformation flows (b), and distortions (c).
• Figure 2: MNIST and EMNIST in the tiny-data regime: the first $1$--$20$ training images per class and the full test set are used. For each classifier, we plot its test accuracy versus the training size $N$ as well as the smallest $N$ needed to reach an accuracy threshold ($90\%$ for MNIST and $75\%$ for EMNIST due to increased difficulty). Our model outperforms all other comparison models for all $N$, requiring the least amount of training data to perform well.
• Figure 3: Omniglot one-shot classification: two sample runs (a) and the error-rate leaderboard (b). The red outline marks one out of $400$ unit tasks, made up of $1$ test and $20$ training images.
• Figure 4: QuickDraw only-one-shot doodle classification: one sample run (a), test accuracies (b), and inter-rater agreement between our model and human performances (c). The sample run exemplifies the training and test images, with the red outline marking a unit task. Test accuracy uses the percentage and the error bar to show the mean and the standard deviation from $5$ independent runs of each model (error bars are omitted for deterministic models). The inter-rater agreement shows the pairwise Fleiss' kappa. Human results are from $5$ healthy subjects ($H_1, \ldots, H_5$) between the ages of $20$ and $29$.
• Figure 5: Archetype generation via $k$-means-style clustering in our learned similarity space.