PaperlayerView exact revision on arXiv

Source-linked AI summary

Teaching neural decoders with language

Xuan Wen, Yonghao Song, Yuhong Zhang, Guoliang Li, Tzyy-Ping Jung, Xiaorong Gao

arXiv:2610.12921v1cs.AI

TL;DR

Class-label supervision leaves relationships among neural-decoding classes implicit, motivating an alternative way to use class semantics. LaS aligns neural features with natural-language class descriptions and predicts by candidate text prototypes; evaluations report gains across decoding tasks, including a 17.15 pp speech gain over CLS for CNN-BiGRU.

  • Problem

    Class-label supervision specifies class identities without explicitly encoding class attributes or semantic relationships, while neural–text alignment raises questions about decoding performance and generalization to unseen classes.

    Source: §1 · pp. 1–2

  • Method

    LaS learns projections that align neural features with class-description text embeddings and predicts by matching each trial to candidate text prototypes.

    Source: §1 · p. 2, §3.2 · p. 3

  • Results

    Paired evaluations report gains across MI, SSVEP, and speech; CNN-BiGRU achieves a 17.15 pp gain over CLS in speech decoding.

    Source: §1 · p. 2, §4.2 · p. 6

  • Takeaways & Limitations

    LaS representations support candidate-set reuse and subject/session adaptation, and LaS enables conditional unseen-frequency decoding in SSVEP.

    Source: §1 · p. 2, §6 · p. 10

    Models are trained independently for each dataset and encoder, so transfer between datasets and recording systems remains to be evaluated.

    Source: §5.4 · p. 10

Abstract

from arXiv · show

Brain-computer interfaces decode neural activity, yet class-label supervision (CLS) leaves relationships among classes implicit. We propose Language as Supervision (LaS), which aligns neural features with natural-language class descriptions through learned projections and predicts classes using candidate text prototypes. We evaluate seven encoder architectures with independently trained models for visual, motor, and speech decoding from scalp and intracranial recordings. LaS improves speech-decoding accuracy by up to 17.15~pp over matched CLS baselines. Its representations support candidate-set reuse and label-supervised subject and session adaptation, while LaS enables conditional unseen-frequency decoding in steady-state visual evoked potential (SSVEP) responses. Correct class-description correspondence supports this generalization, and dual-sided attribution visualization shows neural and textual contributions to the same class-matching score. These findings suggest that text descriptions can serve as general-purpose references for learning and interpretation in neural decoding.

1 INTRODUCTION

Class-label supervision identifies classes without explicitly representing their attributes or semantic relationships, while natural-language descriptions offer an alternative source of semantic information. LaS aligns neural features with text descriptions for learning and candidate prediction, and paired evaluations report gains across motor, visual, and speech decoding.

1 INTRODUCTION

  • One-hot class-label targets specify class identities but do not explicitly encode class attributes or semantic relationships.Natural-language descriptions can provide semantic information that pretrained text models encode.Source: §1 · pp. 1–2
  • LaS aligns projected neural features with class-description embeddings and uses text prototypes as references for candidate prediction.The approach is evaluated alongside established encoder architectures in paired CLS comparisons.Source: Figure 1 · p. 1, §1 · p. 2
  • Paired evaluations across seven encoders report decoding gains in MI, SSVEP, and speech using scalp EEG and intracranial sEEG.LaS also supports candidate configuration, subject/session adaptation, conditional unseen-frequency prediction, and visualization of neural and prompt-word contributions to a matching score.Source: §1 · p. 2

2 RELATED WORK

Prior work connects semantic representations to classification and aligns neural signals with text, using varied training pairings and prediction interfaces. LaS studies class descriptions as both supervision and candidate references across independently trained neural decoders.

2.1 SEMANTIC REPRESENTATIONS FOR CLASSIFICATION

  • Semantic class representations have supported unseen-word fMRI decoding, unseen-class visual prediction, and candidate-description recognition in image and audio models [Elizalde et al., 2023].Related systems also generate classifier weights from descriptions [Ba et al., 2015] or connect frozen image and language models through a Q-Former.Source: §2.1 · p. 2

2.2 LANGUAGE-ALIGNED NEURAL DECODING

  • Neural–text alignment work varies in textual supervision, training-pair units, and prediction use; LaS examines class descriptions as supervision and candidate references across independently trained decoders.Examples include EEG-report alignment for phenotyping [Gijsen & Ritter (2025)], EEG–text prototypes, templated emotion descriptions, and class prototypes with task instructions [Zhang et al., 2026].Source: §2.2 · p. 2

3 METHODS

LaS replaces a task-specific class-label interface with neural–text alignment: each task–encoder model learns projections and predicts among text-described candidates. Its theoretical analysis characterizes how repeated class descriptions and training direction affect contrastive losses and within-class matching.

3.1 PROBLEM SETTING AND CLASS-LABEL SUPERVISION

  • The evaluated tasks include finger MI and SSVEP from scalp EEG and overt Chinese word reading from intracranial sEEG.Speech class descriptions use English translations, and subject-specific sEEG tokenizers feed a backbone shared within each model.Source: §3.1 · p. 2, Figure 2 · p. 3

3.2 LANGUAGE AS SUPERVISION

  • LaS pairs labeled neural trials with fixed class-description prompts, encodes the prompts with a frozen text encoder, and learns projections for neural–text alignment.At prediction, it selects the candidate class with the highest neural–text score; each task–encoder model learns its own alignment space.Source: §3.2 · pp. 3–4, §3.1 · p. 3

3.3 THEORETICAL PROPERTIES

  • With m identical texts per class, N2T has a log m loss floor, while the stated N2T identity assumes balanced class counts and T2N permits unequal counts.Balanced N2T equals class cross-entropy plus a constant; unequal counts reweight present classes.Source: §3.3 · p. 4
  • For classes with multiple trials, T2N implicitly favors equal within-class matching scores when the mean within-class score and other-class scores are fixed.Other feature directions may still differ; for balanced counts nc = m, the listed InfoNCE, Diagonal, and Class-aware SigLIP loss values are log m, m log m − (m − 1) log(m − 1), and 0, respectively.Source: §3.3 · p. 4

4 EXPERIMENTAL EVALUATION

Across MI, SSVEP, and speech decoding, LaS is evaluated against matched class-label supervision, with the largest reported gains in speech. The evaluation also examines candidate-set reuse, label-supervised adaptation, and conditional decoding of unseen SSVEP frequencies.

4.1 EVALUATION SETUP

  • Evaluation uses seven neural encoders across MI, SSVEP, and speech, with task-specific splits and inputs; matched CLS–LaS pairs share encoder architectures and splits.Table 1 reports subject-mean accuracy and mean MI κ; SSVEP uses RW training and IW evaluation.Source: §4.1 · pp. 4–5, Table 1 · p. 5

4.2 DECODING PERFORMANCE

  • 17.15 pp is the largest LaS gain over CLS, achieved by CNN-BiGRU; all four speech encoders show significant paired gains (p < .001).LaS (Du-IN) leads at 60.49% ± 14.30%, a 5.22 pp gain over CLS.Source: Table 1 · p. 5, §4.2 · p. 6
  • LaS (EEG-Conformer) leads MI four-class accuracy at 43.02% ± 9.29%, with mean κ = 0.240 and a 1.36 pp gain over CLS (p < .01).DeepConvNet gains 1.82 pp (p < .05); EEGNet differs nonsignificantly.Source: Table 1 · p. 5, §4.2 · p. 6
  • With RW training and IW evaluation, LaS significantly improves MTSNet at 0.6 s and SSVEPFormer at all three durations; EEGNet differences are nonsignificant and negative at 0.8/1.0 s.In a separate IW-training analysis, LaS leads in accuracy at 0.7–1.0 s and estimated ITR at all five durations, while SSVEPFormer leads in accuracy at 0.6 s.Source: §4.2 · p. 6

4.3 ADAPTATION AND CONDITIONAL UNSEEN-CLASS DECODING

  • LaS adaptation starts from LaS-trained models and uses CLS; options include frozen-feature probes, full fine-tuning, and DTS-ProLIP’s candidate-prototype-restricted projection updates.DTS-ProLIP keeps the encoder and source text prototypes fixed.Source: §4.3 · p. 7
  • For MI session adaptation, full fine-tuning gives the largest mean gains on Thumb/Pinky and Thumb/Index/Pinky; the three-class linear-probe interval crosses zero.Adapted models use eight Base runs and are tested on eight later Fine runs.Source: Figure 4 · p. 6, §4.3 · p. 7
  • For SSVEP subject adaptation, DTS-ProLIP has the highest displayed mean at K = 1, while full fine-tuning leads at K = 2–5; Ensemble TRCA approaches it as K increases.The comparison uses matched calibration budgets and 1-s inputs.Source: Figure 4 · p. 6, §4.3 · p. 7
  • In conditional SSVEP zero-shot decoding, all three LaS encoders numerically outperform FBCCA at 0.6, 0.8, and 1.0 s when testing five held-out frequency classes.The same 35 subjects are used for training and evaluation across eight class splits; chance is 20%.Source: Figure 4 · p. 6, §4.3 · p. 7

5 ANALYSIS AND DISCUSSION

Prompt correspondence matters most for unseen-frequency SSVEP decoding, while attribution visualizes neural and textual contributions to a shared class score. The analysis also finds distinct roles for contrastive directions and identifies transfer and deployment boundaries.

5.1 CLASS CORRESPONDENCE AND LANGUAGE CONFIGURATION

  • P1 leaves closed-set differences nonsignificant but reduces unseen-only SSVEP accuracy to near chance (p < .001).P1 reassigns the unchanged raw text-embedding set, testing class–description correspondence; a text projection can restore seen prototypes under the stated rank and prototype conditions.Source: §5.1 · p. 8
  • With fixed P0 inference prompts, P2 worsens closed-set (p < .001) and unseen-only SSVEP decoding (p < .05), while MI and Speech differences are nonsignificant.With an ensemble of five held-out templates absent from training, P2 improves MI and worsens closed-set SSVEP (both p < .001); Speech and unseen-only SSVEP differences are nonsignificant.Source: §5.1 · p. 8
  • EmbeddingGemma improves closed-set SSVEP over BERT by 2.45 pp (p = .00485), while its 3.40 pp unseen-only gain is nonsignificant (p = .0838).MI and Speech differences are also nonsignificant; freezing versus LoRA adaptation has task- and encoder-dependent effects.Source: §5.1 · p. 8

5.2 INTERPRETING NEURAL–TEXT MATCHES

  • Dual-sided attribution visualizes neural-signal and prompt-word contributions to the same unscaled class-specific cosine score.Examples show task-specific neural and word contributions, including C3 and “thumb”/“flexion” for right-thumb MI, and occipital activity and frequency words for SSVEP.Source: §5.2 · pp. 8–9

5.3 TRAINING OBJECTIVES AND DIRECTIONAL EFFECTS

  • Across tasks, larger batches generally raise InfoNCE and Diagonal SigLIP minimum losses without corresponding accuracy gains; class-aware labels resolve same-class conflicts but do not consistently improve decoding.Class-aware labels also increase positive-pair weights, while eliminating the positive loss floor does not consistently improve accuracy.Source: §5.3 · p. 9
  • Bidirectional InfoNCE outperforms T2N-only in MI and N2T-only in Speech (both p < .001); other directional contrasts are nonsignificant.The comparison covers MI with EEG-Conformer and Speech with Du-IN.Source: §5.3 · p. 9, Figure 8 · p. 9

5.4 LIMITATIONS AND FUTURE WORK

  • Models are trained independently by dataset and encoder, leaving transfer between datasets and recording systems unevaluated.Unseen-class decoding was observed in SSVEP but not Speech; initial SSVEP nonstationarity limits RW-training to IW-evaluation transfer, and online control and long-term stability remain to be evaluated.Source: §5.4 · p. 10

6 CONCLUSION

LaS uses class descriptions as references for supervision and prediction, and the paper presents its experimental protocols across motor, visual, and speech decoding. The appendix specifies shared conventions and task-specific data, model, and evaluation procedures.

6 CONCLUSION

  • LaS uses natural-language class descriptions for supervision and prediction, with neural and textual contributions visualized against a shared class-specific matching score.The conclusion describes independently trained visual, motor, and speech models using scalp and intracranial recordings, plus candidate-set reuse, label-supervised adaptation, and conditional unseen-frequency decoding in SSVEP.Source: §6 · p. 10

A.1 SHARED CONVENTIONS

  • CLS trains a task-specific classification head, whereas LaS pairs neural trials with text using class labels and predicts by matching neural representations to class-text prototypes.For the main comparison, LaS uses frozen EmbeddingGemma with one fixed, semantically matched prompt per class; the neural encoder and both projections remain trainable.Source: §A.1 · p. 16

A.2 MOTOR IMAGERY (MI)

  • The MI protocol uses four finger motor-imagery classes, five within-participant physical-trial folds, and trial-level predictions aggregated from 1-second windows.The appendix specifies EEGNet, DeepConvNet, and EEG-Conformer interfaces, with FBCSP-LDA as a separate conventional MI baseline.Source: §A.2 · pp. 16–17

A.3 STEADY-STATE VISUAL EVOKED POTENTIALS (SSVEP)

  • The SSVEP protocol uses 35 participants and 40 stimulus classes, with cross-subject evaluation and random-window training paired with initial-window evaluation.The compared encoders are EEGNet, SSVEPFormer, and MTSNet; MTSNet includes a frequency branch that zero-pads short windows before representing the complex spectrum.Source: §A.3 · pp. 18–19

A.4 OVERT WORD READING (SPEECH)

  • The Speech protocol uses Chinese word-reading sEEG from 12 participants, with three-second trials and an approximately 80%/10%/10% training, validation, and evaluation split.Four backbones are trained from scratch, and prompts use English translations of the Chinese word targets.Source: §A.4 · pp. 20, 21, 23

A.5 METRICS AND AGGREGATION

The appendix defines task metrics and aggregation, then describes label-supervised adaptation procedures for MI and SSVEP. It also locates additional results and their associated protocols.

A.5 METRICS AND AGGREGATION

  • MI accuracy and Cohen’s κ are aggregated equally across participants, SSVEP accuracy is averaged across evaluation participants, and Speech run results are averaged within participants before equal participant averaging.For MI, LaS averages window logits while CLS and FBCSP-LDA use majority voting; SSVEP estimated ITR is computed within each participant before equal-weight averaging.Source: §A.5 · p. 24

A.6 ADAPTATION ALGORITHMS FOR MI AND SSVEP

  • Adaptation methods initialize from LaS-trained source decoders and update different modules using class-label supervision.Options include frozen-feature probes, learned class weights, projection adaptation with a source-deviation penalty, a candidate-prototype discriminative-subspace constraint, and MLP or full fine-tuning.Source: §A.6 · pp. 24–25

B ADDITIONAL EXPERIMENTAL RESULTS

  • Additional experimental results report subject-level outcomes for Table 1 and Figure 3 and aggregate outcomes for other analyses, using each experiment’s defined evaluation units.The appendix points to dataset protocols, attribution and loss references, and SSVEP window analyses for further detail.Source: B ADDITIONAL EXPERIMENTAL RESULTS · p. 26

B.1 MAIN RESULTS, CANDIDATE-SET REUSE, AND ADAPTATION

This appendix details the protocols for candidate-set reuse, adaptation, and conditional unseen-class decoding, alongside language-side analyses and attribution visualizations. It specifies the evaluation settings and what each analysis compares, but the supplied passages do not report their numerical outcomes.

B.1 MAIN RESULTS, CANDIDATE-SET REUSE, AND ADAPTATION

  • SSVEP initial-window evaluation compares accuracy and estimated ITR across five durations, using IW-trained neural baselines and LaS with inter-trial recombination; all methods are tested on the initial Physical window.This training protocol differs from the random-window-training to initial-window-evaluation comparison in Table 1.Source: §B.1 · p. 26, Table B.1 · p. 26
  • MI candidate-set reuse compares CLS models trained separately for each candidate set with a LaS four-class checkpoint reused for the corresponding candidates.The displayed subsets include classes 2–3, 1–3–4, and 1–2–3, alongside all four classes.Source: §B.1 · p. 29, Table B.10 · p. 30
  • Adaptation analyses cover MI cross-session Base-to-Fine evaluation and SSVEP leave-one-block validation, where K denotes calibration blocks and TRCA requires at least two.The MI comparison includes nine methods across three-class and binary tasks; the SSVEP comparison includes eight neural methods and three TRCA configurations.Source: §B.1 · p. 30, Table B.11 · p. 30, Table B.12 · p. 30

B.2 CONDITIONAL UNSEEN-CLASS DECODING

  • Conditional unseen-class decoding averages five-class accuracy across eight held-out class splits, with each of 40 frequencies contributing 210 Physical evaluation trials.The frequency index is zero-based, with f_i = 8 + 0.2i Hz; neural models use a fixed initial window and five unseen candidates.Source: §B.2 · p. 31, Table B.13 · p. 31, Table B.14 · p. 31

B.3 LANGUAGE-SIDE ANALYSES

  • Language-side analyses compare paired Frozen/LoRA text-encoder runs and evaluate held-out templates individually and as a five-template logit ensemble.The paired runs include BERT-base-uncased and EmbeddingGemma-300M; P0 trains on fixed template 0, while P2 randomly samples five semantic training templates.Source: §B.3 · p. 33, Table B.16 · p. 33, Figure B.1 · p. 33

B.4 ADDITIONAL SIGNAL AND TEXT ATTRIBUTION VISUALIZATIONS

  • Attribution visualizations explain class-matching scores through normalized neural-input groups and signed word contributions, with examples for MI, speech, and SSVEP.MI and speech examples use 120 trials per class for the specified classes; SSVEP examples use six original trials per frequency without inter-trial synthesis or training masking.Source: §B.4 · p. 34, Figure B.2 · p. 35, Figure B.3 · p. 36

B.5 TRAINING OBJECTIVES, DIRECTIONAL ABLATIONS, AND TEMPORAL AGGREGATION

The appendix compares alternative alignment objectives and InfoNCE directions, and examines how MI predictions accumulate across a trial. It also specifies the paired-comparison families and effect-size reporting used for statistical interpretation.

B.5 TRAINING OBJECTIVES, DIRECTIONAL ABLATIONS, AND TEMPORAL AGGREGATION

  • The batch-size grid compares bidirectional InfoNCE, diagonal SigLIP, and class-aware SigLIP across MI, SSVEP, and speech using minimum loss and task-aggregated accuracy.Reference curves use balanced-class formulas.Source: §B.5 · p. 39, Figure B.6 · p. 39
  • Directional ablations compare neural-to-text-only and text-to-neural-only InfoNCE with bidirectional InfoNCE for MI EEG-Conformer and speech Du-IN.The comparison includes paired differences and 95% intervals.Source: §B.5 · p. 39
  • At 5.0 s, logit averaging reaches 43.02%, probability averaging 42.68%, and majority voting 42.38% MI accuracy.The three rules coincide at 33.78% with one 1-s window; subject-mean curves generally rise as overlapping windows accumulate.Source: §B.5 · p. 39, Figure B.7 · p. 40

B.6 PAIRED COMPARISONS AND EFFECT SIZES

  • Table 1 comparisons report LaS-minus-CLS paired differences and paired Cohen’s dz, with tests corrected within task–encoder-specific families.Subjects are the paired units after pooling MI folds or averaging speech seeds within subject; SSVEP uses 35 subjects.Source: §B.6 · p. 41, Table B.17 · p. 41

C.1 NEURAL–TEXT ALIGNMENT: PROPERTIES AND CORRESPONDENCE

LaS aligns neural trials with class prompts, and its repeated-prompt objectives have distinct effects on class scores and within-class responses. The section also characterizes when text prototypes preserve seen-class predictions, how candidate competition affects unseen-class prediction, and how neural and word contributions are attributed to a matching score.

C.1 NEURAL–TEXT ALIGNMENT: PROPERTIES AND CORRESPONDENCE

  • A shared positive denominator preserves LaS’s pairwise class-score rankings, while equal-score sets are hyperplanes in projected neural-feature space and softmax probabilities still use full logits.The encoder from raw signals to neural features may remain nonlinear.Source: §C.1 · pp. 42–43
  • For a fixed neural query, repeating identical class texts adds a parameter-independent constant to the N2T loss without changing class-score gradients; new neural trials add query terms.The decomposition groups repeated text copies by class.Source: §C.1 · p. 43
  • With fixed mean and other-class scores, the T2N class loss is minimized when same-class scores are equal, implicitly aggregating responses along the text-prototype direction when n_c > 1.Other feature directions may still retain trial differences.Source: §C.1 · pp. 43–44
  • A fixed seen-prompt permutation can be compensated to restore seen-class prototypes and pairing scores, but unseen extension additionally depends on the true-description coefficients under the span assumption.Seen-class fitting and unseen-class extension therefore impose distinct correspondence requirements.Source: §C.1 · p. 44
  • In joint seen-and-unseen prediction, a seen-class score can outrank every unseen score even when the true class ranks first within the unseen-only candidate set.Unseen-only prediction uses U; joint prediction uses S ∪ U.Source: §C.1 · p. 44

C.2 REPEATED-PROMPT TRAINING OBJECTIVES

  • Diagonal pair labels treat the n_yi − 1 other same-class trials as negatives, whereas class-aware labels mark all same-class pairs positive.For balanced batches with m copies per class, the other same-class negatives number m − 1.Source: §C.2 · pp. 45–46

C.3 DUAL-SIDED ATTRIBUTION FOR CLASS MATCHING

  • Grouped Integrated Gradients attributes the fixed-text matching score along a path from zero preprocessed input to the full input, with positive and negative values indicating score increases and decreases along that path.The zero baseline has no direct physiological interpretation.Source: §C.3 · p. 47
  • Conditional word attribution zeros one word’s token embeddings and restores that word along a separate path while other prompt content stays fixed; each signed score uses its own baseline.Word scores do not sum to a decomposition of the whole prompt.Source: §C.3 · p. 48

D.1 SSVEP RESPONSES AND SHORT-WINDOW EVALUATION

The section defines the latency-corrected origin and the window intervals used to evaluate short SSVEP segments.

D.1 SSVEP RESPONSES AND SHORT-WINDOW EVALUATION

  • SSVEP windows use a common analysis origin of t0 = 0.14 s; an initial window of length L spans [t0, t0 + L), and offset δ shifts its start.The benchmark recommends an approximately 140-ms shift, though physiological latency may vary among subjects.Source: §D.1 · p. 49

D.2 WINDOW AUGMENTATION AND CLOSED-SET RESULTS

The section describes SSVEP window augmentation and closed-set evaluation. Inter-trial recombination substantially raises initial-window accuracy in the reported BERT-base-uncased condition, while results vary with augmentation and evaluation window.

D.2 WINDOW AUGMENTATION AND CLOSED-SET RESULTS

  • RW samples a valid start within a training trial, while IR concatenates same-class segments from training trials without changing their temporal positions.The RW-trained experiments used approximately 300,000 windows and a shared contiguous temporal mask with ρ = 0.3 for neural models.Source: §D.2 · p. 49
  • The reported IR-volume and mask scans use IW evaluation; the no-IR condition has approximately 6,000 training windows, and mask-scan epoch budgets differ across conditions.The volume scan holds the mask ratio fixed, while the mask scan holds IR volume fixed.Source: §D.2 · p. 50, Table D.5 · p. 51, Table D.6 · p. 51
  • The highest recorded IW accuracy in the mask scan occurs at ρ = 0.3, and RW-trained LaS accuracy is lower on initial Physical windows than on random test windows at all three durations.The mask scan uses approximately 300,000 IR windows; the two evaluation conditions sample different within-trial positions.Source: §D.2 · pp. 49–50, Table D.2 · p. 50, Table D.6 · p. 51
  • 34.7738% without IR rose to 76.3571% with approximately 300,000 recombined windows for BERT-base-uncased initial-window accuracy.At 1.0 s, IR-trained MTSNet LaS with EmbeddingGemma reached 79.50%, compared with the supervised MTSNet result of 78.01% without IR.Source: Table D.3 · p. 50, §D.2 · p. 51, Table D.5 · p. 51

D.3 EARLY-WINDOW BEHAVIOUR AND CONDITIONAL UNSEEN-CLASS DECODING

The section examines how SSVEP accuracy varies with window onset and evaluates conditional unseen-class decoding after IR training. The reported unseen-only results are lower after IR than after RW training at each listed duration.

D.3 EARLY-WINDOW BEHAVIOUR AND CONDITIONAL UNSEEN-CLASS DECODING

  • For one-second Physical windows, the three compared accuracy curves rise through the last measured onset offset, δ = 0.5 s.Each offset uses one window per trial, and Figure D.1 covers six offsets from the valid analysis onset.Source: §D.3 · p. 52, Figure D.1 · p. 52
  • 34.3333%, 65.7857%, and 65.6071%: IR-trained five-class unseen-only accuracy at 0.6, 0.8, and 1.0 s, versus 80.6786%, 87.9762%, and 90.8929% for the RW-trained counterpart.The IR-trained results cover eight class-held-out splits with five unseen candidates per split.Source: §D.3 · pp. 52–53, Table D.7 · p. 52
  • The authors hypothesize that early-window nonstationarity makes additional onset cues harder for LaS to learn and may impair the frequency representation needed for unseen-class transfer.They describe this as a hypothesis consistent with the shared early-window penalty.Source: §D.3 · p. 53

Related papers

Loading 2610.12921v1…