AI Research Atlas
Technical report

Recurrent memory

Prior activity becomes context

Feeding earlier hidden activity back into a network created a learned representation of temporal context.

Jeffrey L. Elman

AI topics

Explore related entries. Larger tags appear on more entries.

The contribution

Elman explored a network whose hidden activity at one time step becomes part of the input for the next. Its internal state therefore reflects both current input and earlier processing. The report studied how such dynamic memory could support sequence prediction and linguistic structure without explicitly storing a fixed window of prior symbols.

What this does not establish

A recurrent state is a learned, compressed context, not a lossless transcript. The surviving 1988 newsletter link contains the report abstract; the author archive supplies the later published version.

Why this date?

The April 1988 CRL newsletter identifies Technical Report 8801. LSTM cites this report. The Cognitive Science article appeared in 1990, and that journal version is the one cited by word2vec and LLaMA.

This entry follows the linked publication. Read the source and date conventions.

Comments

Discuss this research, ask a question, or suggest a correction. Comments appear after the site owner approves them.

Loading comments…

Sign in with ChatGPT to comment

Use your OpenAI account. Published comments show the display name you choose, not your account email.