Amari's recurrent neural network
Red neuronal recurrente de Amari
- Date
- 1972 · confidence medium
- Place
- University of Tokyo, Tokyo, Japan
- Credited to
- Shun'ichi Amari
- Field
- Knowledge & symbols
⚠️ Disputed: the sources disagree.
What was discovered
Amari proposed a neural-network model with recurrent connections able to learn and store sequences of patterns as associative memory, a direct forerunner of Hopfield networks and modern recurrent architectures.
Priority is blurry: Rosenblatt had already described recurrent perceptrons in 1960, and Nakano (1971) and Little (1974) proposed models similar to Amari's; it was the Hopfield network (1982) that popularised the concept. Attributed here to Amari following the Wikidata candidate, with medium date confidence.
What it led to
- Hopfield network (1982)
- Jordan (1986) and Elman (1990) recurrent architectures
- LSTM (Hochreiter and Schmidhuber, 1997)