The Omniglot dataset (Lake et al. 2015) has become a popular dataset for evaluating few-shot learning algorithms. However, there are some variations in the evaluation protocol.
The N-shot K-way learning problem is: given N examples for each of K classes, classify a novel input. For Omniglot, each distinct character is a class. Typical values are K = 5 or 20 and N = 1 or 5.
There are two different ways to obtain few-shot classification problems for testing an algorithm. We will refer to these as "within-alphabet" and "unstructured" evaluation. The difference lies in how a random set of K classes is obtained:
- within-alphabet: Choose an alphabet, then choose K characters from that alphabet (without replacement).
- unstructured: Concatenate the characters of all alphabets, then choose K characters (without replacement). The hierarchy of alphabets and characters is ignored.
Intuitively, we might expect that the unstructured problem is easier, because there is likely to be more variation between alphabets than within alphabets. (This may seem counter-intuitive since characters within an alphabet must be different from one another, whereas characters across alphabets may be identical. However, a character in one alphabet can have at most one such near-identical match in another alphabet.)
The original Omniglot github repo uses within-alphabet evaluation.
It defines 20 runs, each of which comprises 20 training images and 20 testing images from the same alphabet (2 runs for each of the 10 evaluation alphabets; see lake-evaluation/run_alphabets.txt for the correspondence).
We used within-alphabet evaluation in our NIPS 2016 paper, although we used many random trials instead of 20 runs.
In contrast, several recent papers have used unstructured evaluation. For example, in the ProtoNets source code, they load a list of all characters and then take a random subset of these characters.
Here we train a simple siamese network and compare the numbers obtained using different evaluation methods.
We report results in terms of classification error (lower is better; chance is 1 - 1/K).
The results support the hypothesis that within-alphabet problems are much harder than unstructured problems.
| unstructured | within-alphabet | |
|---|---|---|
| 20-way 1-shot | 10.7% | 24.2% |
| 20-way 5-shot | 3.6% | 12.8% |
| 5-way 1-shot | 3.5% | 9.5% |
| 5-way 5-shot | 1.1% | 4.2% |
Hence, while recent papers may seem to be approaching saturation of the Omniglot task, we should remember that there is still room for improvement in the within-alphabet task.
Details for this experiment are as follows (the results can be replicated by running python train.py --data_dir=data/ --download).
The embedding network uses more or less the same architecture as the Matching Nets paper (Vinyals et al. 2016).
We input 24px images to a 4-layer conv-net, where each hidden layer has 64 channels.
There are relu, batch-norm and max-pooling (stride 2) operations between each linear layer (but not at the output).
The 64-D embedding vectors are compared using a cosine distance.
During training, each batch contains 16 few-shot problems, each of which contains K = 5 classes with 1 training image and 1 testing image.
We then use the cross-entropy-of-softmax loss with one-hot labels.
To make predictions, we use a 1-NN (nearest neighbour) classifier.
The situation is further complicated by the use of different splits.
- The official Omniglot repo defines a
backgroundset of 30 alphabets and anevaluationset of 20 alphabets. - Koch et al. (2015) refer to a 40/10 split of alphabets.
- The Prototypical Networks repo contains train/val/test splits which partition the 50 alphabets into 33/5/12 alphabets respectively.
These are known as the Vinyals splits, presumably because they were used in the Matching Nets paper.
The
testset is a subset of the originalevaluationset. We list the alphabets in this repository atsplits/vinyals/.