Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,14 @@
|
|
| 1 |
---
|
| 2 |
license: cc-by-4.0
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: cc-by-4.0
|
| 3 |
---
|
| 4 |
+
|
| 5 |
+
This is a subset of the [Multilingual Spoken Word Corpus](https://huggingface.co/datasets/MLCommons/ml_spoken_words) dataset, which is built specifically for the Few-shot Class-incremental Learning ([FSCIL](https://github.com/xyutao/fscil)) task.
|
| 6 |
+
|
| 7 |
+
A total of 15 languages are chosen, split into 5 base languages (English, German, Catalan, French, Kinyarwanda) and 10 incrementally learned languages (Persian, Spanish, Russian, Welsh, Italian, Basque, Polish, Esparanto, Portuguese, Dutch).
|
| 8 |
+
The FSCIL task entails first training a model using abundant training data on words from the 5 base languages, then in subsequent incremental sessions the model must learn new words from an incremental language with few training examples for each, while retaining knowledge of all prior learned words.
|
| 9 |
+
|
| 10 |
+
Each of the 5 base languages consists of 20 classes, with 500/100/100 samples for train/val/test splits each.
|
| 11 |
+
Each of the 10 incremental languages consists of 10 classes, each with 200 available samples. From these, a small number (e.g., 10) will be chosen for few-shot training, and 100 other samples are chosen for testing.
|
| 12 |
+
Thus, the model first has a knowledge base of 100 words from the base classes, which expands to 200 words by the end of all incremental sessions.
|
| 13 |
+
|
| 14 |
+
All audio is in opus (48kHz) format.
|