K Means Assign — Machine Learning/Classical Models

K Means Assign

Machine Learning / Classical Models

Assigns the input to the nearest of K fitted cluster centroids, and reports how far away it was:

d²_c = Σ_j (u_j − C_cj)², idx = argmin_c d²_c + base, d² = min_c d²_c

This is the inference half of k-means – KMeans.predict with cluster_centers_ pasted in. The distance output is what makes it usable as a monitor rather than only a classifier: a sample far from every centroid still gets an index, and only the distance says the model has never seen anything like it.

Ports

• u – the sample, a column [d,1]. d must equal the number of COLUMNS of Centroids, and it is checked rather than broadcast.
• idx – a scalar [1,1]: which centroid won, plus Index Base.
• d2 – a scalar [1,1]: the squared distance to that centroid. See Notes – the square root is deliberately not taken here.

Parameters

• Centroids – C, a [K,d] matrix with one centroid per row. That is KMeans.cluster_centers_'s own orientation, so it pastes in without transposing. K is the number of clusters and d the feature count.
• Index Base – what the first cluster is called:
  • Zero-based (sklearn) – the first cluster is 0. The default, and what KMeans.predict returns.
  • One-based (MATLAB) – the first cluster is 1.
• 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. Every centroid is unrolled and baked in as literals at full setprecision(17); there is no tunable parameter, because a fitted centroid is not something to retune on the target.

The three HDL targets are ordinary Q16.16 fixed point and are offered for synthesis: squares, sums and comparisons only.

Simulink bridge

None. Simulink has no k-means block – clustering lives in the Statistics and Machine Learning Toolbox as MATLAB functions, not as library blocks, so there is nothing with parameters for this to map onto. 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.

Notes

• Algebraic and stateless: the outputs depend only on the current input, so the block cannot break an algebraic loop.
• Nonlinear and discontinuous, and deliberately carries no state space.
• Why the distance is squared. Ranking by squared distance and by distance give the same winner – the square root is monotone – so omitting it changes no decision and keeps every backend exact and the HDL targets synthesizable. If you want the Euclidean distance, put a Sqrt block on d2 and pay for the root only where you need it.
• Ties go to the lowest index, the same rule Argmax Decision uses, mirrored for a minimum: the scan replaces its running best only on a strictly smaller distance. Two blocks that both reduce a vector to a decision must agree about this or a graph using both contradicts itself.
• At a near-tie the HDL targets can assign differently, for the reason Argmax Decision's description spells out: they compare Q16.16 values where the in-app run compares doubles, so two centroids nearly equidistant can rank the other way and the index output then differs by a whole cluster. That is inherent to a discrete decision in fixed point, not a codegen defect.