Glossary

Definitions for terms used across EMMA. Define an acronym on first use on every page that uses it; this page is the canonical reference.

Signal and RF

RF

Radio-frequency. The band of the electromagnetic spectrum used by wireless communications, radar, and spectrum sensing. EMMA evaluates models on RF data.

I/Q

In-phase and quadrature. The two-component complex baseband representation of a signal: channel 0 is the in-phase component, channel 1 is the quadrature component. Stored as float32 (real and imaginary parts) for autograd compatibility. EMMA’s canonical input.

multi-antenna I/Q

I/Q captured across more than one antenna element simultaneously, preserving the phase difference between antennas. Required for AoA, DoA, and beam tasks. The property spectrograms destroy and CSI discards.

phase coherence

A fixed, known phase relationship between antennas at the receiver. The physically necessary ingredient for direction-finding from an array, and EMMA’s input-layer moat.

AoA

Angle of arrival. The direction from which a signal reaches a receiver array, estimated from inter-antenna phase differences. EMMA’s flagship regression task (E-LOC-AOA).

DoA

Direction of arrival. Synonym for AoA in the array-processing literature; used when both azimuth and elevation matter.

CSI

Channel state information. A matrix or tensor describing the wireless channel between transmitter and receiver. Operated on directly by the CSI camp (LWM, DeepMIMO). EMMA offers CSI only as a bridge task, not the primary input.

ULA

Uniform linear array. Antenna elements equally spaced along a line. Used for azimuth-only tasks.

URA

Uniform rectangular array. Antenna elements on a 2D grid. Used when elevation and azimuth are both needed.

SNR

Signal-to-noise ratio. The ratio of signal power to noise power, in decibels. An OOD axis for several tasks.

SEI

Specific emitter identification. Identifying an individual device from its unintentional RF signature. Synonymous here with RF fingerprinting.

RF fingerprinting

Identifying a specific device (not just its waveform) from hardware imperfections such as CFO, IQ imbalance, and phase noise. EMMA’s E-ID-FP task, scored leave-one-unit-out.

CFO

Carrier frequency offset. A hardware-induced offset between the intended and actual carrier frequency; a fingerprint feature.

LoS / NLoS

Line-of-sight and non-line-of-sight. Whether a direct path exists between transmitter and receiver. EMMA’s E-LOC-LOS task.

micro-Doppler

The fine Doppler signature produced by moving parts of a target (rotor blades, limbs, vehicle wheels). The basis of E-ID-UAVDOP.

AMC

Automatic modulation classification. Classifying the modulation scheme of a signal (QPSK, 16QAM, and so on). The RadioML task. EMMA includes it only as a continuity column.

SigMF

Signal Metadata Format. An open standard for recording signal metadata alongside I/Q samples. EMMA’s interchange format with rfgen.

Benchmark and methodology

OOD

Out-of-distribution. Evaluation on data that differs from training along a controlled axis (environment, device, frequency band, SNR). The score, not an appendix.

leave-one-environment-out

An OOD protocol where the training split spans a set of channel environments and evaluation holds out one environment at a time, averaging the metric over held-out environments.

leave-one-unit-out

An OOD protocol for device fingerprinting where all captures of one device are held out at test time, isolating device identity from channel and capture-session confounds.

sim-to-real gap

The difference between a model’s score on synthetic data and its score on a real-capture OOD subset, reported in percentage points. A large gap means the model fit the generator rather than learned signal physics.

OOD-avg

EMMA’s headline aggregate score: the normalized mean of a model’s per-task scores, each computed under its OOD axis. Lower-is-better metrics are inverted before normalization.

content hash

A SHA-256 digest of the (rfgen commit, config, seed) triple that determines a scene. Makes a dataset re-derivable and a release auditable.

frozen backbone

A pretrained foundation model whose weights are not updated during task evaluation. EMMA’s SUPERB-style upstream contract.

readout head

A lightweight, task-specific network attached to a frozen backbone to produce task outputs (an angle, a class distribution, a caption).

SUPERB

Speech processing Universal PERformance Benchmark. The structural template EMMA follows: one frozen backbone, lightweight readouts, one aggregated score.

Metrics

MAE

Mean angular error. The mean absolute difference between predicted and true angle, in degrees. The metric for E-LOC-AOA.

AUC

Area under the ROC (receiver operating characteristic) curve. A threshold-free ranking metric for binary detection tasks.

AUROC

Area under the ROC curve, used for anomaly and novelty detection.

EER

Equal error rate. The point at which the false-accept and false-reject rates of a verifier cross. Used for fingerprinting.

NMSE

Normalized mean squared error. Used for channel estimation and CSI feedback.

SI-SDR

Scale-invariant signal-to-distortion ratio. A standard source-separation metric. The metric for E-SC-SEP.

BLEU / METEOR / CIDEr

Captioning metrics from image-captioning and machine-translation literature: BLEU (Bilingual Evaluation Understudy), METEOR (Metric for Evaluation of Translation with Explicit ORdering), and CIDEr (Consensus-based Image Description Evaluation). Used for E-S2T-CAP.

exact-match / F1

Span question-answering metrics from SQuAD. Used for E-S2T-QA.

Projects and precedents

rfgen

The rf-data-generation framework. EMMA’s primary data source.

RadioML

DeepSig’s modulation-classification datasets (RML2016, RML2018). The field’s de-facto AMC benchmark: saturated, single-antenna, and its publisher states the data has known errata and is not used in its products.

DeepMIMO

A channel-dataset generator by Alkhateeb. Operates at CSI level. The data source for the LWM lineage.

LWM

Large Wireless Model. A CSI-native foundation model by the DeepMIMO group.

IQFM

A raw complex I/Q foundation model by Mashaal and Abou-Zeid. The closest existing analog to the model EMMA evaluates.

PReD-Bench

A multi-task electromagnetic benchmark (2026) that uses spectrograms and constellations, discarding phase. The cautionary tale for operator-as-evaluator.

Colosseum / POWDER / COSMOS

Public RF testbeds (Northeastern and NSF PAWR). The real-capture OOD sources for EMMA’s sim-to-real column.

WILDS

A benchmark of in-the-wild distribution shifts (Koh et al., 2021). A generalization-first precedent for EMMA’s OOD framing.