Low sample efficiency has plagued deep learning researchers in their attempts to advance machine perception. Although a human child can confidently recognize a fluffy panda after a single visual encounter, a computer vision model might need thousands of training images to do the same. This has made few-shot learning an attractive research field.
“Few-shot learning is the problem of learning new tasks from little amounts of labeled data. This topic has gained tremendous interest in the past few years, with several new methods being proposed each month,” Google Brain Group in Montréal Lead Hugo Larochelle said in his keynote at the recent RE•WORK Deep Learning Summit in Montréal.
Larochelle suggested the research community take a step back and take stock of what’s already been achieved.
Recent deep learning research has proved the ability of deep neural networks to extract complex statistics and learn high-level features from huge amounts of data. In the case of limited datasets, Larochelle suggested researchers explore few-shot learning, citing progress in computer vision as an example. “Since I last spoke, Mini-ImageNet – 5 shot (Few-Shot Image Classification on Mini-ImageNet – 5-Shot learning) has made significant improvements. This is through both various method developments in recent months alongside an influx of extremely talented individuals choosing to focus their work and research in this topic area.”
Larochelle said studies also indicate the possibility of achieving high accuracy even without labeled data. “In fact, the accuracy of results in few-shot learning, both with and without labeled data, is very high.” He said researchers have used prototypical networks for the problem of few-shot classification, where a classifier has to generalize to new classes that were not seen in the training set while only a small number of examples of each new class are available. Another group of researchers from the University of California, Berkeley, and OpenAI proposed an algorithm for meta-learning that is model-agnostic. It shows good results on few-shot regression and accelerated fine-tuning for policy gradient reinforcement learning with neural network policies.
The improvements in few-shot learning approaches have lead Larochelle and others in the research community to ask whether better benchmarks are needed. Said Larochelle, “One question we can ask: Are the labels of the support set actually useful? I don’t really need those labels because I can infer them from the support set… So we propose a class semantics consistency criterion (CSCC).”
Larochelle said Google Brain has also proposed a benchmark called “Meta-Dataset” and presented the following guidelines for determining what a better few-shot learning benchmark would look like:
- Learning across many tasks requires learning over many datasets
- Vary the number of classes per episode
- Vary the number of examples per episode
- Vary the relative frequency of each class per episode
Larochelle is confident few-shot learning performance will continue to improve and become more effective. He closed his talk with three takeaways:
- We can achieve surprisingly high accuracy on current few-shot learning benchmarks, without using labels.
- The most popular few-shot learning methods aren’t robust to training on heterogeneous sets of tasks.
- None of these few-shot learning methods dominate in all settings (e.g. of number of shots).
Journalist: Fangyu Cai | Editor: Michael Sarazen