Confidence Gate — Machine Learning/Postprocessing

Confidence Gate

Machine Learning / Postprocessing

Reports the winning class only when it is worth acting on:

max = u_argmax(u), valid = 1 when max ≥ Threshold and 0 otherwise, and idx = argmax(u) + base when valid, the configured Fallback Index when not.

This is what lets a controller ignore a low-confidence classification instead of switching on it. Put it after Softmax, where the threshold reads as a probability, or after a bare Dense Layer, where it reads as a logit margin. It is Argmax Decision plus one comparison, and it breaks ties the same way – the lowest index wins.

Ports

• u – the scores, a column [m,1] with m ≥ 1 – logits, probabilities or any comparable quantity.
• idx – a scalar [1,1]: the winning position plus Index Base when the gate opens, and Fallback Index exactly as configured when it does not. It is carried as a real number like every other signal in the model; it is always a whole one.
• valid – a scalar [1,1], 1 or 0: whether the winning score cleared the threshold. Emitted as a number rather than a boolean, so no backend needs a type the others lack.
• max – a scalar [1,1]: the winning entry's value, in whatever units u carried. Reported whether or not the gate opened, so a downstream monitor can see how close a rejected sample came.

Parameters

• Threshold – a scalar. The gate opens when the winning score is at or above it; the comparison is on the raw score, in the units u carries, so after a softmax it is a probability in [0,1] and after a dense layer it is a logit. Default 0.5.
• Fallback Index – a scalar, reported on idx when the gate stays shut. It is passed through exactly as configured and is not shifted by Index Base – it names no class, so there is nothing to shift. Default −1, which no zero- or one-based class can take and so reads as "no decision"; set it to a real class index instead if the graph should fall back to a default class.
• Index Base – what the first element is called:
  • Zero-based (PyTorch, numpy) – the first class is 0. The default, and what a model exported from a Python framework expects.
  • One-based (MATLAB) – the first class is 1, for a graph whose downstream lookup or switch is written in MATLAB's convention.
  It shifts the reported winning index and nothing else – the comparison, the tie rule, the threshold test, valid, max and the fallback are identical either way.
• 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. The threshold, the fallback and the index base are baked into the body at export time; there is no tunable parameter, so retuning the operating point means re-exporting.

The three HDL targets are genuine Q16.16 fixed point – the block only compares and copies, with no transcendental in it – but see the Notes on what that costs at a near-tie and at the threshold itself. A Fallback Index outside roughly ±32768 cannot be represented there and will saturate; the software targets carry it exactly.

Simulink bridge

None. The closest block on this installation is dspstat3/Maximum, whose Value and Index mode does reproduce the reduction and even carries an indexBase parameter – but it has no threshold, no fallback and no validity output, and its dialog (probed in R2026a, not read from documentation) offers no parameter any of the three could map onto. Gating a decision on its own confidence is the whole of this block and none of it would cross. The bridge reports the block 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

• Algebraic and stateless: the outputs depend only on the current input, so the block cannot break an algebraic loop. In particular the gate has no hysteresis – a score hovering at the threshold makes valid chatter, exactly as a comparison should. Follow it with Debounced Decision or a Relay on max if the application needs the decision to stick.
• Nonlinear and discontinuous, and deliberately carries no state space. Both idx and valid step rather than vary, so no A/B/C/D describes the block and model reduction correctly refuses it.
• Ties go to the lowest index, the same rule Argmax Decision and K Means Assign use. The scan replaces its running best only on a strictly greater value.
• At a near-tie, or at the threshold, 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 two scores closer than that can rank the other way round, and a score within a quantum of the threshold can fall on the other side of the gate. idx then differs by a whole index and valid by 1, rather than by a rounding error. This is inherent to reducing a continuous quantity to a discrete decision in fixed point, not a codegen defect: place the threshold away from where the scores pile up, or make the decision downstream in floating point, if your application must agree bit for bit with the simulation.