The forerunner of NTL was the L₀ project, launched in 1990 by Profs. Jerome Feldman (UC Berkeley computer science and then-director of ICSI), and George Lakoff (UC Berkeley linguistics). The goal of the project was to build systems that could learn a fragment of a natural language from a set of picture-sentence pairs, where the sentence describes something about the picture. The group was named L₀ because they expected to learn approximately 0% of any given language. Lokendra Shastri joined the L₀ project in 1993 when he moved to ICSI from the University of Pennsylvania.

A number of UCB computer science dissertations grew out of the L₀ effort:

• Terry Regier built a system capable of learning the meaning of spatial relations such as "on" and "under" from pictorial examples, each labeled with a spatial relation between two of the objects in the scene.
• Andreas Stolcke demonstrated how a grammar can be induced from examples using Bayesian model merging, in which hypotheses about a grammar are updated based upon labeled examples and are merged to minimize the complexity of the grammar.
• David Bailey's VerbLearn system uses model merging to learn the different senses of action verbs such as "push" and "slide" from labeled examples of structured event descriptions.
• Srini Narayanan's KARMA system uses metaphorical mapping and the simulation of events/actions to draw metaphorically entailed inferences from preparsed text input.

In 1997, the name of the group was changed to the Neural Theory of Language (NTL) to reflect a shift in focus from pure language learning to computational models of all aspects of language (including learning, performance, and understanding) with an emphasis on biologically plausible approaches. The group has also become more interdisciplinary, with students from both computer science and linguistics.