Monty.
Copy Markdown
View Source
Attention that learns where looking has paid off: Monty's
attention/attention_system.py::DefaultAttentionSystem.
Each step it turns the sensor's proposed regions into a sparse grid of weights, lets the previous grid decay, drops the voxels that have faded to nothing, and merges the new grid into the old one. The merged grid is then what the goal filter reads, so a goal is weighed against everything attention has seen, not just against this step's proposal.
The pipeline, in Monty's order:
proposed = regions → voxels → one pooled weight per voxel
step: proposed → telemetry
decay(grid) → expire → merge(expired, proposed) → telemetry
goal_filter(grid, goals) → goalsexpire/1 is the part that makes the grid a memory rather than an
accumulator: a weight whose magnitude has decayed below 1e-6 is dropped, so a
place that has not been looked at for long enough stops being a voxel at all.
The collaborators — weight_pooler, decay, merge, goal_filter — are
functions, which is this port's form of Monty's one-method protocols; new/1
turns a decay value into one for you. telemetry is a value it threads, because
telemetry is the one collaborator with state of its own.
Summary
Types
A decay: a grid in, the same grid one step closer to zero out.
A goal filter: what attention knows, and the goals, in; goals to pursue, out.
A merge: two grids in, one grid out.
A weight pooler: many weights about one voxel in, one weight out.
Functions
The grid without the voxels whose weight has faded below the tolerance.
A system that has learned nothing yet.
Forget every voxel and everything the telemetry recorded.
The grid attention has built, and what the telemetry recorded.
Act on one step: learn where the sensors looked, and say which goals survive.
The weight magnitude below which a voxel is dropped from the grid.
Types
@type decay() :: (Monty.Attention.VoxelGrid.t() -> Monty.Attention.VoxelGrid.t())
A decay: a grid in, the same grid one step closer to zero out.
@type goal_filter() :: (Monty.Attention.VoxelGrid.t(), [Monty.Cmp.Goal.t()] -> [Monty.Cmp.Goal.t()])
A goal filter: what attention knows, and the goals, in; goals to pursue, out.
@type merge() :: (Monty.Attention.VoxelGrid.t(), Monty.Attention.VoxelGrid.t() -> Monty.Attention.VoxelGrid.t())
A merge: two grids in, one grid out.
@type t() :: %Monty.Attention.AttentionSystem.DefaultAttentionSystem{ decay: decay(), goal_filter: goal_filter(), grid: Monty.Attention.VoxelGrid.t(), merge: merge(), telemetry: term(), voxel_size: float(), weight_pooler: weight_pooler() }
A weight pooler: many weights about one voxel in, one weight out.
Functions
@spec expire(Monty.Attention.VoxelGrid.t()) :: Monty.Attention.VoxelGrid.t()
The grid without the voxels whose weight has faded below the tolerance.
A weight whose magnitude is under 1e-6 says nothing worth remembering, and
keeping it would keep the grid growing forever.
A system that has learned nothing yet.
Defaults are Monty's: 5 cm voxels, negative_priority_max_pool, linear decay at
Monty's default rate, Union, HardGoalFilter, and telemetry that records
nothing.
Forget every voxel and everything the telemetry recorded.
The grid attention has built, and what the telemetry recorded.
Act on one step: learn where the sensors looked, and say which goals survive.
Answers the system to use next step plus the goals worth pursuing. The goals go through the goal filter after the merge, so they are judged against the memory including this step's proposals.
@spec weight_expiration_tolerance() :: float()
The weight magnitude below which a voxel is dropped from the grid.