Embedding Lookup — Machine Learning/Neural Networks

Embedding Lookup

Machine Learning / Neural Networks

Selects one row of a trained embedding table by index: y = table[k, :], delivered as a column [d,1].

This is what turns a token, a category or a mode number into the dense vector a network was trained on – the front end of any model whose input is not a measurement. The index arrives as an ordinary signal and is rounded to the nearest whole number, so 2.0 and 1.9997 select the same row.

Ports

• k – the index, a scalar [1,1]. Interpreted relative to Index Base.
• Output – y, a column [d,1], where d is the NUMBER OF COLUMNS of Embedding Table. The selected row transposed into a column; all zeros when the index is out of range and Out Of Range is All Zeros.

Parameters

• Embedding Table – the trained table, [V,d]: one row per index, which is the orientation torch.nn.Embedding.weight and a Keras Embedding layer both store, so weights paste in without transposing. Its shape is the only size source – V is the number of indices the block knows and d is the width of the output, and there is no second config that could disagree with it.
• Index Base – what the first row is called:
  • Zero-based (PyTorch, numpy) – index 0 selects the first row. The default.
  • One-based (MATLAB) – index 1 selects the first row.
• Out Of Range – what an index outside the V rows does:
  • All Zeros – the output is entirely zero. The default, and the one that lets a downstream block notice that nothing was selected. It is also what a padding or unknown token usually means.
  • Clamp To Range – below the first row selects the first, above the last selects the last.
  An index outside the table is not exotic: an unmapped token or a glitched mode number produces one on a live system, and the two settings send it to opposite places.
• 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 table, the base and the out-of-range rule are structural and baked in; there is no tunable parameter.

No backend converts the index to an integer, and no backend indexes the table with a signal. The lookup is emitted as one half-open comparison per row – (k − base) − v in −0.5, +0.5) – guarding a copy of that row's d constants. That is the same rule One Hot Encoder uses, it needs no integer type, and it leaves nothing in the emitted code that could read outside the table.

The three HDL targets are ordinary Q16.16 fixed point and are offered for synthesis: one subtraction, V comparisons and d assignments, with no multiplier and no transcendental. The cost is V×d constants in the emitted body rather than one indexed read – the honest price of a lookup that carries no address arithmetic, at the table sizes a deployed model uses.

Simulink bridge

None. Simulink's Deep Learning blocks take a trained network object, which no parameter mapping can carry, and Direct Lookup Table (n-D) is not this block: it takes an index per dimension and returns the single element they address, not a whole row. 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 output depends only on the current index, so the block cannot break an algebraic loop.
• Nonlinear and discontinuous, and deliberately carries no state space.
• The rounding boundary is a discontinuity, and the HDL targets sit one quantum away from the reference. They compare Q16.16 values (one quantum ≈ 1.5e-5) where the in-app run compares doubles, so an index within that distance of a ±0.5 boundary can round the other way and select a different row – and a different row is a whole different vector, not a rounding error. Feed this block a genuinely integral index – which is what a token is – and the question never arises. The same caveat governs One Hot Encoder and Argmax Decision.
• Ordinary numbers, not a fitted object: the table is config, so nothing is loaded at run time and every target exports without a framework.