iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Geoffrey Hinton’s 2021 hunch about what might come next for AI was GLOM, a proposal for helping neural networks represent how parts fit into wholes. It was an idea for a possible architecture, not a working AI system: Hinton’s paper explicitly says it “does not describe a working system.”
What is GLOM?
GLOM is Hinton’s proposal for representing part-whole hierarchies in a neural network. The idea is that a network could build a structured interpretation of an image or other input rather than treating it only as a collection of separate features. Hinton described GLOM as “a single idea about representation” that draws on advances from several research groups.
The name refers to a conceptual design, not a released product or a demonstrated, general-purpose AI architecture. In his 2021 feature for MIT Technology Review, Siobhan Roberts characterized it as an intuition; Hinton called it “vaporware” at that point.
Free tools Windows power users keep installed
One-click scans. No signup required.
How does GLOM represent parts and wholes?
GLOM’s central mechanism is “islands of identical vectors.” In simplified terms, units in a neural network produce vector representations—lists of numbers encoding an interpretation. Nearby units that arrive at sufficiently similar representations can form an island, standing for a node in a hierarchy.
#1 Best Overall
- A smaller island could represent a part, such as an eye or a wheel.
- Another island could represent a larger structure made from parts, such as a face or a car.
- Islands at different levels could together encode a parse tree: a structured account of how the input’s components relate.
In this design, the network’s architecture stays fixed while its pattern of agreeing vector groups changes with the input. One image might produce a hierarchy for a face; another might produce a different hierarchy. Roberts’s “islands of agreement” phrasing is a helpful analogy for the proposed mechanism, not evidence that the network works through social-style agreement.
What problem was Hinton trying to solve?
Hinton’s proposal addressed two related challenges in visual perception, as described in the 2021 feature. First, recognizing a whole scene requires identifying objects and understanding their natural parts. Second, recognizing an object from a new viewpoint requires a representation that is not tied too narrowly to one familiar angle.
Rank #2
Hinton hoped that a system able to represent these relationships explicitly might make neural-network representations easier to interpret, including in systems built for vision or language. He also saw this kind of idea as one possible ingredient in more flexible, human-like problem solving. Those were goals and hopes, not established abilities of GLOM.
Was GLOM a working AI system?
No. Hinton’s primary paper, “How to represent part-whole hierarchies in a neural network”, submitted to arXiv on February 25, 2021, states that it “does not describe a working system.” The paper lays out a conceptual proposal rather than reporting a finished architecture with demonstrated performance.
At the time Roberts’s feature appeared on April 16, 2021, Google colleagues were investigating preliminary, highly supervised experiments involving simple arrangements of ellipses. The feature reported that there was not yet enough evidence to assess the idea’s real significance. The ellipse setup was not a performance benchmark, and the available accounts do not establish that GLOM had solved general vision, achieved state-of-the-art results, or been deployed.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What can be said about GLOM’s status now?
The cited paper and feature establish what Hinton proposed and what was known in 2021; they do not establish what research or implementations followed. Without newer evidence, it would be inaccurate to describe GLOM as either a current working system or an abandoned idea. The sound conclusion from these sources is narrower: GLOM was an ambitious hypothesis about representation whose practical significance was still uncertain when the feature was published.
Quick Recap
Best Value
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
