sample post · a layout template · text adapted from Ying Nian Wu's comment on the paper · clip computed for this sample
Deqian Kong

understanding how networks represent, reason, and learn

Planning follows a straight path

November 2026 · note

Place cells in the hippocampus fire when an animal is at a particular place. Instead of asking what a single cell does, take the whole population at position x as one vector, h(x, t), where t sets the spatial scale.The scale grows like \sqrt{t}, which mirrors the dorsoventral axis of the hippocampus. We ask one thing of this vector: its inner products should be the transition probabilities of a random walk,

\langle\, h(x,t),\; h(y,t) \,\rangle \;=\; q(y \mid x,\, t) (1)

where q(y \mid x, t) is the probability of reaching y from x in t steps, normalized so that q(x \mid x, t) = 1. If the elements of h are also required to be non-negative, each one becomes a localized place field, and together the fields tile the space.

Multi-scale place fields and planned paths in open fields and mazes

Figure 1. Place fields at several time scales, and the paths planned from them in open fields and mazes.

Why the path is straight

Every h(x, t) has unit norm, so (1) gives

\Vert\, h(x,t) - h(y,t) \,\Vert^2 \;=\; 2 - 2\, q(y \mid x,\, t) (2)

Climbing q toward a goal y is therefore the same as descending the Euclidean distance to h(y, t). In the maze the route bends around walls. In the embedding it heads straight for the goal.

Many scales

One scale is not enough. Near the goal a small t gives a sharp signal, and far from it only a large t reaches at all. The scales are tied together by

P_{2t} \;=\; P_t^{\,2} (3)

which is the recursion of dynamic programming.No successful trajectory is needed to learn this. The walk is random. At each step the planner uses the scale with the largest gradient.

Video 1. The field q(y \mid x, t) for a fixed goal as t grows, then the path found by the planner below.

# Setup omitted: A is the adjacency matrix of the free cells of the maze,
# deg its row sums, and P = I/2 + A/(2 deg) the lazy random walk on them.

def q(Pt):
    """Normalized transition probability, so that q(x | x, t) = 1."""
    K = Pt / deg[None, :]        # symmetric kernel of the reversible walk
    d = np.sqrt(np.diag(K))
    return K / np.outer(d, d)

# Scales t = 1, 2, 4, ...: P_2t = P_t squared.
scales, Pt = [], P.copy()
for k in range(12):
    scales.append(q(Pt))
    Pt = Pt @ Pt

def plan(x, y):
    """Greedy ascent on q(y | x, t), at the scale with the largest gain."""
    path = [x]
    while x != y:
        nbrs = np.nonzero(A[x])[0]
        gain = np.array([[Q[n, y] - Q[x, y] for n in nbrs] for Q in scales])
        x = nbrs[gain.max(0).argmax()]
        path.append(x)
    return path
symbolmeaning
x,\; ypositions
tnumber of steps of the walk
h(x, t)population vector of place cells
q(y \mid x, t)normalized transition probability

Table 1. Notation.

Reference

  1. M. Zhao, D. Xu, D. Kong, W.-H. Zhang, Y. N. Wu. Place Cells as Multi-Scale Position Embeddings: Random Walk Transition Kernels for Path Planning. NeurIPS 2025.