Debounced Decision — Machine Learning/Postprocessing

Debounced Decision

Machine Learning / Postprocessing

Holds the class it is currently reporting until N consecutive samples agree on a different one. The index passes through unchanged in value – what changes is when it is allowed to move.

This is what stops a classifier chattering a controller between modes on a single noisy sample. Put it after Argmax Decision, whose index it takes and whose index it returns.

A lone disagreeing sample cannot move the output at all, and it throws away whatever evidence had been collected. That is the LATCH behaviour; if you want a lone sample to be outvoted rather than to reset the count, use Temporal Majority Vote, which is the same decision-stabiliser in its smoothing form.

Ports

• k – the raw class index, a scalar [1,1], as Argmax Decision reports it. Class i is the one whose window [i−0.5, i+0.5) the index falls in, after Index Base has been subtracted.
• cls – a scalar [1,1]: the class being held, reported in the same base as the input.
• count – a scalar [1,1]: how many consecutive samples have now agreed on the pending candidate. It is 0 whenever the input agrees with the class already held, and it returns to 0 on the sample that switches. Watch it to see a switch coming.

Parameters

• Number Of Classes – K, a whole number ≥ 1. The classes are the K windows starting at the index base; anything outside them is not a class.
• Index Base – what the first class is called:
  • Zero-based (PyTorch, numpy) – the first class is 0. The default.
  • One-based (MATLAB) – the first class is 1.
  It shifts both the input window and the reported index by the same amount, so a graph that is consistent about the base cannot notice it.
• Confirmations – N, a whole number ≥ 1: how many consecutive samples must agree on a new class before the output moves. At N = 1 the block is a pass-through (every valid sample switches immediately), which is the documented degenerate case rather than a misconfiguration.
• Initial Class – the class position held before any evidence arrives, counted from 0 whatever the base. It must be less than K.
• Sampling Time (s) – zero or less inherits the solver's rate; a positive value runs the block at that period.

Code export

All ten targets: Python, MATLAB, Java, Rust, C, C++, VHDL, Verilog, SystemVerilog and PLC Structured Text. K, the base, N and the initial class are baked into the body at export time; there is no tunable parameter, because every one of them changes which comparisons are emitted rather than merely what they compare against.

The three HDL targets are genuine synthesizable Q16.16, not the simulation-only real path: the whole block is comparisons, constant assignments and one increment, with no division and no transcendental. The class windows are compared in fixed point directly, so nothing converts through an integer type.

Simulink bridge

None. Simulink has no decision debouncer: Detect Change reports that a signal has moved, and the Discrete library's delays and counters would have to be assembled into this behaviour rather than mapping onto it, so there is no parameter set this block could cross through. The bridge reports it rather than dropping it silently, and it has no parity testbench, which is the documented consequence of Support::None rather than a gap. Code export verification still covers it across all ten languages.

Notes

• Stateful, and inherently discrete – the run counter advances once per sample, so the block declares setDiscreteOnlyBlock and takes its rate from its own Sampling Time (s). It carries no state space and model reduction correctly refuses it: the output steps rather than varies.
• An index outside every class window is not evidence. It leaves the held class, the pending candidate and the count exactly as they were. A dropout or an unmapped category therefore cannot discard evidence already collected, and cannot contribute any.
• The switch takes effect on the confirming sample – the Nth agreeing sample already reports the new class, rather than the one after it.
• At a window boundary the HDL targets can decide differently. They compare Q16.16 values, one quantum of which is about 1.5e-5, where the in-app run compares doubles – so an index within that distance of a ±0.5 boundary can fall in the neighbouring class, and the decision then differs by a whole class rather than by a rounding error. This is inherent to reducing a continuous quantity to a discrete class in fixed point, and it is the same caveat Argmax Decision and Embedding Lookup carry. Feed this block a genuinely integral index – which is what Argmax Decision produces – and no sample is ever near a boundary.