Neural Network Plans and Imagines Like a Brain: Study in Nature Machine Intelligence

A team led by Lin published a brain-inspired generative model in Nature Machine Intelligence on July 21, 2026. The model plans actions and solves tasks using cognitive maps, and it learns only through local synaptic plasticity.

What the model is built on

The model from Lin and colleagues combines three mechanisms, each borrowed from neuroscience. At its core are cognitive maps: internal representations of the environment that let the system "chart a route" to a goal before taking any real action.

• Publication: July 21, 2026, Nature Machine Intelligence (DOI 10.1038/s42256-026-01254-4)
• Authors: Lin and colleagues (Lin et al.)
• Three pillars of the model: cognitive maps, stochastic computation, compositional coding
• Training: local synaptic plasticity only, without global backpropagation
• Two functions of intelligence: planning and problem solving

Stochastic computation means the model works through random sampling (neural sampling). Instead of producing one rigid answer, it explores possible futures. Compositional coding lets it assemble new scenes and plans from familiar elements, much as a person imagines a situation they have never been in.

How this differs from conventional neural networks

The model from Lin and colleagues differs from most neural networks in how it learns. Today's systems train with backpropagation, which requires global access to every weight in the network and is biologically implausible. Here, each connection updates according to a local rule, based on signals from its nearest neighbors, as is presumed to happen in a living brain.

Local learning brings the architecture closer to biology and is potentially more economical for neuromorphic hardware, where global operations are expensive. According to the authors, the model delivers two key traits of intelligence at once, traits that conventional generative systems imitate only in part.

"The model provides two key traits of intelligence: planning and problem solving," from the paper by

Lin et al. in Nature Machine Intelligence.

What is goal-directed imagination?

Goal-directed imagination is the model's ability to mentally "play out" paths to a given goal. Rather than simply reacting to the current input, the system samples possible action sequences from the cognitive map and picks those that lead to the goal.

This approach brings machine planning closer to how the brain operates: an imagined route first, then a real step. It is the combination of three elements (a cognitive map, stochastic sampling and local learning) that sets this work apart from purely engineering planners, which make no claim to biological plausibility and typically rely on global learning.

What this means

The work by Lin and colleagues is a step toward neural networks that plan and imagine on brain-like principles rather than only recognizing patterns. Local learning combined with stochastic inference fits well with energy-efficient neuromorphic chips and with robotics, where the two stated functions, planning and problem solving, are especially valuable.