MEX Operator
Mathematical Methods AIMech:MEXOperator DRAFT
Aliases: MEX (SimNets)
A scalar aggregation operator that returns log(mean(exp(xi times input)))/xi for a nonzero real parameter xi, with its continuous zero-parameter extension equal to the arithmetic mean.
Identity Notes & Concept Scope
Compared all 442 pinned AIO and 618 native IDs, names, aliases and scoped synonyms; no normalized identity match. Ignored- and hidden-inclusive fixed-string searches of the recovery, older discovery and original Documents checkouts returned 125, 0, 80 lines. All distinct match contents were inspected. Semantic matching covered 43 AIO and 31 native definitions, with every matched definition read. 5 full annotated AIO node/edge/generated contexts and 2 full native records were inspected. MEX denotes the scalar log-mean-exp operator used by Cohen and Shashua, separate from the complete SimNet architecture and a spatial MEX layer with offsets and sharing. The search also found MATLAB mex syntax and unrelated identifier-prefix registries in ignored dependencies; those are different meanings. The alias is qualified as MEX (SimNets). A softmax probability vector and the unnormalized log-sum-exp within lifted structured loss are related computations, not this operator identity.
Claims & Grounded Evidence
4 Claims
Mechanism
#normalized-scalar-and-limits
For a nonempty finite real input list, MEX uses the logarithm of the mean exponential divided by xi. It equals scaled log-sum-exp minus log(N)/xi, so dropping the mean normalization changes the finite value. The zero-parameter extension is the arithmetic mean, and the positive and negative infinite-parameter limits are maximum and minimum. Its output is a scalar, not a probability vector. Differentiating with respect to the inputs at fixed nonzero xi yields softmax weights of xi times those inputs; this derivative relationship does not equate the two outputs.
Mechanism
#nested-collapse-conditions
Applying the operator within equal-sized groups and again across their outputs with the same parameter gives the same value as applying it to all entries at once. Equal-size groups preserve the common mean normalization. Unweighted nesting over unequal group sizes generally assigns different weights to flattened entries and need not collapse; using different parameters across stages also generally breaks this identity. The operator is invariant to permutation of its inputs and shifts by the same amount when a common constant is added to every input.
Scope
#relu-and-layer-scope
The paper builds MEX layers from the operator applied to local input blocks with additive learned offsets, optional constants and sharing choices. Ordinary max pooling is obtained as the positive infinite-parameter limit over the selected block. Recovering scalar ReLU additionally requires a zero input and zero offsets; finite MEX of an input and zero is a smooth aggregate and can be negative. The mathematical operator is therefore distinct from a particular layer layout and from the complete SimNet architecture, which also uses similarity operations.
Limitation
#combined-network-evidence-boundary
The evaluated shallow SimNet uses one MEX stage with parameter one and a final spatial mean stage, corresponding to a zero-parameter limit. These stages do not satisfy the common-parameter collapse condition. Reported accuracy and parameter counts concern the whole trained classifier with its similarity functions, initialization, preprocessing and optimization. They do not establish an isolated MEX-versus-other-operator accuracy, speed or memory advantage.
⚠️ Curated Limitations & Negative Results
- MEX here is the SimNets operator, not MATLAB mex tooling or similarly named identifier registries. The record describes a scalar operation, not every MEX layer or a complete network.
- The direct exponential formula can overflow in finite-precision arithmetic; constructed examples use a shifted log-sum-exp calculation. Mathematical limits do not prescribe a numerically stable zero-parameter implementation.
- Equal group size and a common parameter are material to the unweighted nested-collapse identity. Finite smoothing, derivative weights and exact limiting maxima remain distinct.
- The complete original 19-page primary version is reused with its original access timestamp and hash. Author code, later Deep SimNets full text and model training were not inspected; DRAFT remains.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:MathematicalFunction
Mathematical Function
|
broad_match | MEX is a parameterized scalar rule on a finite nonempty list, within the broader pinned mathematical-function class. |
AIO:SoftmaxFunction
Softmax Function
|
related_match | At fixed nonzero xi the input gradient of MEX is softmax of xi-scaled inputs. This derivative relation is useful but does not make the scalar value a probability vector. |
AIO:ReLUFunction
ReLU Function
|
related_match | Adding a zero input and taking the positive infinite-parameter limit recovers scalar ReLU; finite MEX is not equivalent to the pinned function. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
cohen2014
|
Nadav Cohen, Amnon Shashua
|
primary paper
full_text
|
ID: arXiv:1410.0781v3
sha256:91119773f55d…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-13T12:48:39Z
Updated: 2026-10-13T12:48:39Z
Source File:
data/curated/MEXOperator.yaml