A Look Back at DeepMind's Grid-LSTM

prajdabre · x · 2026-07-19

The author revisits the era of diverse neural network architectures that preceded Transformer, using Grid-LSTM (2015, DeepMind) as a primary example.

They explain that the core of Grid-LSTM was that a cell no longer had a single output and memory state, but multiple states, allowing the same set of inputs to be observed and interact along multiple axes. The author likens this approach to a "multi-perspective data representation."

They further connect this to later concepts like multi-head attention and "parameter sharing across time and depth axes," suggesting these ideas look like precursors to today's universal/looping transformer. The final verdict: this work seemed outrageous at the time, but in hindsight, it was incredibly ahead of its time.

Original post →

More from Research

Research channel →