Supervised Contrastive Learning Framework for Electroencephalography-based Air-writing Recognition

Sup ervised Con trastiv e Learning F ramew ork for Electro encephalograph y-based Air-writing Recognition Anan t Jain 1 ⋆ [0000 − 0002 − 7131 − 8310] and A yush T ripathi 2 ⋆ [0000 − 0002 − 7944 − 2260] 1 Sc ho ol of T echnology , IFIM College, Bangalore, India anant.jain@ifim.edu.in 2 Departmen t of Electrical Engineering, Indian Institute of T ec hnology Indore, India ayush.tripathi@iiti.ac.in Abstract. Electro encephalograph y (EEG) - based air-writing recogni- tion offers a human-computer in teraction paradigm b y deco ding neural activit y asso ciated with handwriting mov ements. Despite its p otential, reliable EEG-based air-writing recognition remains challenging due to lo w signal-to-noise ratio and pronounced inter-sub ject v ariability . In this study , we examine the use of sup ervised contrastiv e learning to improv e represen tation learning for EEG-based air-writing recognition. The anal- ysis is conducted on prepro cessed EEG signals and indep enden t comp o- nen t analysis (ICA)–derived neural comp onents obtained from fiv e par- ticipan ts, with trials segmented from − 1 to 2 s relative to mo vemen t on- set. EEGNet and DeepConvNet architectures are ev aluated under b oth con ven tional cross-entrop y training and a supervised con trastive learning framew ork using a sub ject-dep endent fiv e-fold cross-v alidation scheme. The results indicate that sup ervised con trastive learning consisten tly im- pro ves classification accuracy across arc hitectures and feature represen- tations. F or prepro cessed EEG signals, the mean accuracy increases from 33.45% to 43.77% and from 29.14% to 38.06% with EEGNet and Deep- Con vNet, respectively . Using ICA components, higher mean accuracies of 49.21% and 43.32% are achiev ed with EEGNet and DeepCon vNet, re- sp ectiv ely . These results suggest that the sup ervised contrastiv e learning framew ork offers an efficien t extension to existing EEG-based air-writing recognition approaches. Keyw ords: Electroencephalography (EEG) · Air-writing · Sup ervised con trastive learning · EEGNet · Indep enden t comp onen t analysis. 1 In tro duction 1.1 Bac kground and Related W ork Air-writing recognition seeks to identify characters written in free space by un- constrained finger or wrist motions [2, 3]. This offers a natural, touchless input ⋆ Authors contributed equally to this work. 2 A. Jain and A. T ripathi framew ork for emerging h uman–computer in teraction (HCI) technologies [12]. In con trast to gesture-based interfaces, air-writing does not require users to adopt artificial gestures, as c haracters can be formed using con v entional writ- ing motions. Air-writing is particularly suitable for w earable computing [13], augmen ted and virtual realit y environmen ts [17], and assistive tec hnologies [23]. Most recent air-writing recognition systems utilize non-neural sensing modali- ties suc h as inertial measurement units (IMUs) [19, 20], surface electromy ogra- ph y (EMG) [21, 22], or computer vision-based tec hniques [10, 24]. Despite their robust p erformance, these mo dalities dep end on p eripheral sensors that are sen- sitiv e to device-sp ecific c haracteristics and exhibit limited generalization across users. These limitations motiv ate the exploration of neural sensing mo dalities that capture scalp brain signals corresponding to motor execution. Electro encephalograph y (EEG) pro vides a non-inv asiv e method for recording cerebral activity asso ciated with motor planning and execution [6]. EEG-based BCI systems are p opular due to lo w cost, portability , and high temporal reso- lution [7]. EEG-based brain–computer interface (BCI) systems ha ve b een used for motion recognition in v arious motor-execution tasks, suc h as biceps curl [15], grasp-and-lift [8], and turn detection [1]. Pei et al. [14] prop osed an EEG- based handwritten letters recognition system using a con volutional neural net- w ork (CNN) deco der. Crell et al. [4] used EEG-based letter kinematics deco ding to recognize 10 letters. V arious EEG-based features w ere explored for the recog- nition of 21 letters by T ripathi et al [18]. How ever, most EEG-based air-writing recognition metho ds rely on traditional cross-entrop y–based sup ervised learn- ing. It primarily fo cuses on optimizing decision b oundaries and ma y struggle to learn robust and discriminativ e represen tations. Recently , Supervised contrastiv e learning (SPL) [9] has emerged as an efficient alternativ e, sp ecifically promoting in tra-class compactness and in ter-class separability within the learned em b ed- ding space. The study explores the effectiveness of SPL in air-writing recognition using EEG signals. 1.2 Ob jectives and Contributions This study aims to integrate sup ervised con trastive learning and EEG-based air-writing recognition. The ob jectives of the study are to apply a supervised con trastive learning framework to scalp-recorded EEG signals and ev aluate its efficacy in learning discriminative neural representations for air-written c haracter classification. The key contributions of the study are summ arized as follo ws: 1. Dev eloping a sup ervised con trastiv e learning framework for EEG-based air- writing recognition, facilitating structured represen tation learning from scalp EEG signals. 2. In tegration of deep-learning architectures and performance ev aluation using pro cessed EEG and ICA-derived feature represen tations. 3. Comparativ e analysis b et ween sup ervised con trastive learning and conv en- tional cross-entrop y–based training in a multi-sub ject EEG air-writing set- ting. Title Suppressed Due to Excessive Length 3 This study expands contrastiv e representation learning to EEG-based air- writing recognition and offers a compact y et efficient methodological enhance- men t for neural HCI systems. 2 Materials and Metho ds 2.1 Data Description In this study , the publicly av ailable Neur oAiR dataset [18] is utilized. The dataset comprises EEG recordings from health y right-handed participan ts p erforming an air-writing task. Data corresponding to five sub jects are analyzed in this study . During the recording sessions, participants were seated comfortably and instructed to remain relaxed, with their righ t elb ow resting on a table to minimize upp er-arm mov ement and reduce motion-related artifacts. A custom graphical user interface developed with the Tkinter module in Python was used to presen t visual cues for uppercase English letters. Participan ts interacted with the in ter- face using their non-dominan t hand, while writing the prompted characters in free space using their index finger. EEG signals w ere recorded using a 31-channel LiveAmp EEG system (Brain Pro ducts Gm bH, Germany), with electro des placed according to the in terna- tional 10–20 system using an EasyCap electrode cap. The signals were recorded with a sampling rate of 500 Hz. Electrode imp edance w as maintained b elo w 20 k Ω throughout the recording sessions to ensure adequate signal quality . A t the start of eac h trial, the participant initiated the task b y pressing the space bar, after which a single c haracter w as presen ted on the screen. The participant was instructed to write the display ed c haracter in free space using the index finger. The trial w as ended with the press of the space bar again. Eac h participan t completed 100 repetitions of the full set of 26 upp ercase English letters, whic h yielded a total of 2600 samples p er sub ject. Eac h letter was recorded individu- ally , and short breaks were provided after ev ery five sets of recordings to reduce fatigue. 2.2 EEG Preprocessing EEG prepro cessing w as performed using EEGLAB [5] to olb o x in MA TLAB soft- w are. The prepro cessing enhanced the EEG signal qualit y b y reducing artifacts and extracting neural feature representations. Raw EEG signals were common- a verage-referenced. Subsequen tly , the signals were bandpass-filtered b et w een 0.5 and 45 Hz using zero-phase, non-causal finite-impulse-resp onse (FIR) filters to remo ve slow drifts and high-frequency noise. F urther, independent comp onen t analysis (ICA) was applied to decompose the EEG signals into statistically inde- p enden t source comp onents. Under the ICA mo del, the m ultichannel EEG signal is represented as a linear combination of indep enden t comp onents, enabling the separation of neural sources from non-neural artifacts such as ey e blinks and m uscle activity . Artifact-related comp onents were iden tified using the ICLab el 4 A. Jain and A. T ripathi algorithm implemented in EEGLAB. The comp onen ts corresponding to the o c- ular or muscular artifacts, with a confidence score exceeding 0.8, were remov ed. Clean EEG signals were then reconstructed by recombining the remaining com- p onen ts using a modified mixing matrix. In this study , only tw o feature representations were retained for subsequen t analysis: Pr epr o c esse d EEG signals , corresp onding to artifact-free EEG data, and ICA c omp onent time series , represen ting spatially filtered neural source activi- ties. Both representations were segmented into fixed-length trials corresp onding to the air-writing task and normalized using z-score normalization prior to mo del training. 2.3 Sup ervised Contrastiv e Learning F ramework T o enhance the discriminativ e p o wer of EEG-based air-writing represen tations, this study employs a sup ervise d c ontr astive le arning (SCL) framew ork. Let a mini-batc h of N labeled EEG trials b e denoted as { ( x i , y i ) } N i =1 , where x i represen ts an EEG trial corresp onding to an air-written character and y i ∈ { 1 , . . . , C } indicates its class label. Eac h input trial is mapp ed into a latent represen tation via an enco der netw ork, r i = Enc ( x i ) , (1) where Enc ( · ) corresponds to a deep neural net work enco der. The latent repre- sen tation r i is then normalized and passed through a pro jection head to obtain z i = Pro j ( r i ) , (2) where Pro j ( · ) denotes a nonlinear mapping implemen ted using a fully connected la yer. The sup ervised con trastive loss promotes samples b elonging to the same class to form compact clusters in the em b edding space while enforcing separation b et w een samples from different classes. F or a giv en anchor sample i , let P ( i ) denote the set of indices corresp onding to samples in the mini-batch that share the same class lab el as i , and let A ( i ) denote the set of all samples in the batch except i . The supervised con trastive loss is defined as L sup = N X i =1 − 1 | P ( i ) | X p ∈ P ( i ) log exp( z i · z p /τ ) P a ∈ A ( i ) exp( z i · z a /τ ) , (3) where τ is a temp erature scaling parameter and ( · ) denotes the inner product b et w een normalized embedding vectors. The minimization of the loss function explicitly promotes intr a-class simi- larity and inter-class dissimilarity within the learned embedding space. In con- trast to the conv entional cross-entrop y loss, whic h fo cuses on optimizing decision b oundaries, the sup ervised con trastive loss directly shap es the geometry of the laten t space. Hence, it improv es represen tation robustness in the presence of in ter-trial and inter-sub ject v ariability commonly observed in EEG signals. Title Suppressed Due to Excessive Length 5 The ov erall learning strategy follows a two-stage tr aining pr oto c ol . In the first stage, the enco der and pro jection head are jointly optimized using the sup ervised con trastive loss to learn discriminativ e EEG representations. In the second stage, the pro jection head is discarded, and a linear classification head is trained on the frozen enco der outputs using cross-entrop y loss. This strategy ensures that, at inference time, the mo del complexity remains iden tical to that of a standard classification netw ork while leveraging contrastiv ely learned representations. 2.4 Mo del Architectures In the present study , tw o established conv olutional neural netw ork architec- tures— EEGNet [11] and De epConvNet [16] —are utilized for EEG-based air- writing recognition. These models were selected due to their demonstrated ef- fectiv eness in EEG deco ding tasks. Both architectures are ev aluated with and without the sup ervise d c ontr astive le arning (SCL) fr amework across all five par- ticipan ts. EEGNet EEGNet is a compact con volutional neural netw ork specifically de- signed for EEG-based brain–computer interface applications. The architecture seeks to extract interpretable spatial–temporal features from multic hannel EEG signals with constrained training data. EEGNet consists of tw o primary stages. In the first stage, temporal conv olutions are applied to the input EEG signal using t wo-dimensional conv olutional filters to learn frequency-selectiv e repre- sen tations. This is follow ed b y a depth wise conv olution op eration that acts as a spatial filter across EEG c hannels, enabling the mo del to learn c hannel-wise spatial patterns asso ciated with neural activity . Batch normalization and the ex- p onen tial linear unit (ELU) activ ation function are applied to stabilize training and introduce nonlinearity , resp ectiv ely . A verage p ooling and drop out lay ers are utilized to reduce dimensionality and av oid ov erfitting. In the second stage, sep- arable con volutions are employ ed to enhance the represen tation learning while main taining a minimal parameter coun t. The output feature maps are flattened and passed to a fully connected classification lay er with softmax activ ation in the cross-en tropy–based training setup. In the sup ervised con trastive learning framew ork, the softmax classifier is replaced by a pro jection head during the represen tation learning stage, while the enco der remains unc hanged. DeepCon vNet DeepCon vNet is a con volutional neural netw ork arc hitecture comprising a sequence of conv olutional and po oling blocks that hierarchically extract temporal and spatial features from EEG signals. The first block of Deep- Con vNet consists of consecutiv e conv olutional lay ers that perform temp oral fil- tering, follo wed by spatial filtering across EEG c hannels. Subsequent blo c ks con- sist of conv olutional la yers with increasing num b ers of filters, eac h follo wed by batc h normalization, ELU activ ation, and max-p o oling op erations. This hierar- c hical structure enables the model to learn progressiv ely abstract representations 6 A. Jain and A. T ripathi T able 1. EEGNet Mo del Architecture La yer Configuration Input EEG signal ( 31 × 1500 ) Con v2D 8 filters, kernel (1 × 64) Batc h Normalization – Depth wise Conv2D k ernel (31 × 1) , depth multiplier = 2 Batc h Normalization – A ctiv ation ELU A v erage Pooling (1 × 4) Drop out 0.5 Separable Conv2D k ernel (1 × 16) Batc h Normalization – A ctiv ation ELU A v erage Pooling (1 × 8) Flatten – Dense (CE setup) 26 units, softmax Dense (SCL setup) Pro jection head (during Stage 1) of EEG activity related to air-writing mov ements. In the conv entional classifica- tion setting, the output of the final conv olutional blo c k is flattened and fed into a fully connected softmax la yer trained using cross-entrop y loss. In the sup ervised con trastive learning configuration, the DeepCon vNet enco der is coupled with a pro jection head during the first training stage, as in the EEGNet-based SCL setup. F or both EEGNet and DeepConvNet, the supervised con trastive learning framew ork is implemented b y treating the conv olutional backbone as an enc o der network . During the first stage of training, the enco der is enhanced with a pro jec- tion head and optimized using sup ervised contrastiv e loss to learn discriminative laten t representations. The pro jection head comprises of a single fully connected la yer with 128 neurons and ReLU activ ation. The enco der and pro jection head parameters are join tly optimized b y minimizing the sup ervised contrastiv e loss using the Adam optimizer. During the classification stage, the pro jection head is discarded and substituted b y a fully connected output la yer consisting 26 neu- rons with softmax activ ation. A drop out rate of 0.5 is applied to the classifier to reduce o verfitting, and the classification net work is trained using the standard cross-en tropy loss. T o ev aluate the impact of con trastive representation learning, eac h mo del is trained and ev aluated in t wo configurations: 1. Standard sup ervised learning, using cross-en tropy loss alone, and 2. Sup ervised contrastiv e learning–based training, following the tw o-stage pro- to col. All exp erimen ts are uniformly conducted across the fiv e participants, facili- tating a direct comparison of architectures and training strategies while isolating the effect of sup ervised con trastive learning on EEG-based air-writing recogni- tion. Title Suppressed Due to Excessive Length 7 T able 2. DeepConvNet Mo del Architecture La yer Configuration Input EEG signal ( 31 × 1500 ) Con v2D 25 filters, kernel (1 × 5) Con v2D 25 filters, kernel (31 × 1) Batc h Normalization – A ctiv ation ELU Max Pooling (1 × 2) Con v2D 50 filters, kernel (1 × 5) Batc h Normalization – A ctiv ation ELU Max Pooling (1 × 2) Con v2D 100 filters, kernel (1 × 5) Batc h Normalization – A ctiv ation ELU Max Pooling (1 × 2) Con v2D 200 filters, kernel (1 × 5) Batc h Normalization – A ctiv ation ELU Max Pooling (1 × 2) Flatten – Dense (CE setup) 26 units, softmax Dense (SCL setup) Pro jection head (during Stage 1) 3 Results and Discussion 3.1 Exp erimen tal Details In this study , experiments w ere conducted using pr epr o c esse d EEG signals and indep endent c omp onent analysis (ICA) c omp onent time series as input features for EEG-based air-writing recognition. F or eac h trial, EEG segmen ts were ex- tracted from a temp oral window spanning − 1 s to 2 s relative to the onset of the air-writing task. This window was selected to capture both pre-mo v ement neu- ral activit y and execution-related dynamics. Since EEG signals were recorded at a sampling rate of 500 Hz , each trial segmen t consisted of 1500 time samples. F or shorter-duration trials, zero-padding was applied to ensure a fixed-length represen tation across all samples. The resulting input feature matrices w ere of dimension κ × 1500 , where κ = 31 for b oth prepro cessed EEG signals and ICA comp onen t time series. All input data were normalized using z-score normaliza- tion to achiev e a mean of 0 and a v ariance of 1. A user-dep endent 5-fold cr oss-validation technique w as employ ed to ev aluate the model’s p erformance. F or eac h participan t, the dataset was split in to five m utually exclusive folds. F our folds are used for training, while one fold is re- serv ed for testing in eac h iteration. This pro cedure was rep eated until each fold functioned as the test set once. Within eac h training fold, a subset of the data w as allo cated as a v alidation set for early stopping. The p erformance metrics 8 A. Jain and A. T ripathi T able 3. Mean accuracy score across the participants using pre-pro cessed EEG and ICA comp onents time series data in cross-entrop y loss and sup ervised contrastiv e loss settings. Pre-pro cessed EEG participan ts Cross-En tropy Loss Supervised Contrastiv e Loss EEGNet DeepCon vNet EEGNet DeepCon vNet Sub01 33.58 33.85 43.88 39.58 Sub02 30.62 24.00 38.73 34.54 Sub03 35.31 28.23 55.88 44.31 Sub04 42.69 36.85 50.92 43.65 Sub05 25.04 22.77 29.42 28.23 A verage 33.45 29.14 43.77 38.06 ICA Comp onen ts participan ts Cross-En tropy Loss Supervised Contrastiv e Loss EEGNet DeepCon vNet EEGNet DeepCon vNet Sub01 35.04 35.62 49.65 44.77 Sub02 35.04 29.31 44.46 38.88 Sub03 41.12 30.62 59.04 50.35 Sub04 43.62 42.12 57.88 49.85 Sub05 27.65 24.73 35.00 32.73 A verage 36.49 32.48 49.21 43.32 w ere a veraged across the folds to obtain sub ject-sp ecific results. All deep learn- ing mo dels were trained with a mini-b atch size of 32 . Mo del optimization was p erformed using the Adam optimizer with early stopping configured to a pa- tience of 20 on v alidation accuracy . The final reported results correspond to the a verage p erformance across all five participants, ensuring a uniform ev aluation of the prop osed and baseline mo dels. 3.2 Results T able 3 summarizes the me an classific ation ac cur acy obtained across five partic- ipan ts using pr epr o c esse d EEG signals and ICA c omp onent time series , ev alu- ated under cr oss-entr opy (CE) and sup ervise d c ontr astive le arning (SCL) train- ing paradigms with EEGNet and DeepConvNet architectures. F or prepro cessed EEG inputs, mo dels trained using sup ervised contrastiv e learning consistently outp erform their cross-entrop y-trained coun terparts across all participants and arc hitectures. A v eraged across participan ts, EEGNet trained with CE ac hieves a mean accuracy of 33.45%, which increases to 43.77% when trained using the SCL framework. A similar trend is observ ed for DeepCon vNet, with mean ac- curacy increasing from 29.14% (CE) to 38.06% (SCL). Mean accuracy analysis sho ws that the SCL framework yields significan t improv emen ts in classification p erformance. F or Sub03, an increase in accuracy from 35.31% (CE) to 55.88% (SCL) is observed with the EEGNet deco der. This shows the efficiency of sup er- vised contrastiv e representation learning for class separability . Overall, EEGNet Title Suppressed Due to Excessive Length 9 consisten tly outperforms DeepCon vNet on preprocessed EEG data across b oth training strategies. Impro vemen t in classification p erformance is observed when ICA comp onent time series were utilized as input features across training paradigms and arc hi- tectures. With cross-en tropy training, the mean accuracy increases to 36.49% for EEGNet and 32.48% for DeepConvNet. F urther, the sup ervised contrastiv e learning framework improv ed the accuracy to 49.21% and 43.32% for EEGNet and DeepCon vNet, respectively . F or ICA-based features, the impro vemen ts in p erformance using the SCL framew ork are significantly greater than those with prepro cessed EEG signals. The highest a verage accuracy of 49.21% is achiev ed with EEGNet and the SCL framework using ICA-features, an improv ement of ≈ 12.7% o ver the cross-en tropy baseline. SCL framework consistently out- p erforms the traditional cross-en tropy-based training framew ork. F urthermore, ICA-deriv ed features outp erform prepro cessed EEG signals, which shows the significance of spatially decomp osed neural representations for EEG-based air- writing recognition. Mean accuracy scores demonstrate that integrating sup er- vised contrastiv e learning with deep learning classifiers yields b etter discrimina- tiv e laten t represen tations for EEG-based air-writing recognition. 4 Discussion The exp erimen tal results demonstrate the efficiency of the sup ervised con trastive learning framew ork for EEG-based air-writing recognition in comparison to cross- en tropy-based training. The improv ements are observed in b oth prepro cessed EEG signals and ICA comp onen t-based features, using b oth the EEGNet and DeepCon vNet classifiers. ICA-based features achiev e higher classification accu- racy than prepro cessed EEG signals. This indicates that indep endent comp onen t represen tations pro vide b etter discriminative input for learning neural em b ed- dings. Across both training configurations, EEGNet outp erforms DeepConvNet, highligh ting the imp ortance of compact, EEG-sp ecific architectural induc tiv e bi- ases. Overall, the classification analysis sho ws the robustness and efficiency of the sup ervised contrastiv e learning framework for EEG-based air-writing recog- nition. 5 Conclusion This study introduced the sup ervised contrastiv e learning framew ork for EEG- based air-writing recognition. The preprocessed EEG signals and ICA-deriv ed neural comp onents w ere used as input features. The proposed framework in- tegrates con trastive representation learning with EEGNet and DeepCon vNet classifiers to impro ve classification p erformance. Exp erimen tal results demon- strate the sup eriority of sup ervised contrastiv e learning ov er conv entional cross- en tropy-based training. A significan t p erformance impro vemen t while using ICA 10 A. Jain and A. T ripathi comp onen t features. EEGNet outp erforms DeepConvNet, showing the effective- ness of compact EEG-sp ecific architectures. Overall, the framework provides an effectiv e extension to existing EEG-based air-writing recognition frameworks. References 1. Anand, P ., Jain, A., Muth ukrishnan, S.P ., Bhasin, S., Roy , S., Kumar, L.: EEG- based 90-Degree T urn In tention Detection for Brain-Computer Interface. arXiv preprin t arXiv:2410.11339 (2024) 2. Chen, M., AlRegib, G., Juang, B.H.: Air-writing recognition—Part I: Modeling and recognition of characters, words, and connecting motions. IEEE T ransactions on Human-Machine Systems 46 (3), 403–413 (2015) 3. Chen, M., AlRegib, G., Juang, B.H.: Air-writing recognition—P art I I: Detection and recognition of writing activit y in con tinuous stream of motion data. IEEE T ransactions on Human-Machine Systems 46 (3), 436–444 (2015) 4. Crell, M.R., Müller-Putz, G.R.: Handwritten character classification from EEG through con tinuous kinematic decoding. Computers in Biology and Medicine 182 , 109132 (2024) 5. Delorme, A., Makeig, S.: EEGLAB: an open source to olb o x for analysis of single- trial EEG dynamics including indep endent comp onen t analysis. Journal of neuro- science metho ds 134 (1), 9–21 (2004) 6. F alk, T.H., Guger, C., V olosyak, I.: Brain–computer interfaces: recent adv ances, c hallenges, and future directions. Handb ook of Human-Machine Systems pp. 11–22 (2023) 7. Jain, A., Kumar, L.: PreMovNet: Premov ement EEG-based hand kinematics esti- mation for grasp-and-lift task. IEEE Sensors Letters 6 (7), 1–4 (2022) 8. Jain, A., Kumar, L.: ESI-GAL: EEG source imaging-based tra jectory estimation for grasp and lift task. Computers in Biology and Medicine 186 , 109608 (2025) 9. Khosla, P ., T eterwak, P ., W ang, C., Sarna, A., Tian, Y., Isola, P ., Masc hinot, A., Liu, C., Krishnan, D.: Supervised contrastiv e learning. A dv ances in neural information pro cessing systems 33 , 18661–18673 (2020) 10. Kim, U.H., Hwang, Y., Lee, S.K., Kim, J.H.: W riting in the air: Unconstrained text recognition from finger mo vemen t using spatio-temporal conv olution. IEEE T ransactions on Artificial Intelligence 4 (6), 1386–1398 (2022) 11. La whern, V.J., Solon, A.J., W ayto wich, N.R., Gordon, S.M., Hung, C.P ., Lance, B.J.: EEGNet: a compact con volutional neural netw ork for EEG-based brain– computer interfaces. Journal of neural engineering 15 (5), 056013 (2018) 12. Mitra, S., Ac harya, T.: Gesture recognition: A survey . IEEE T ransactions on Sys- tems, Man, and Cybernetics, Part C (Applications and Reviews) 37 (3), 311–324 (2007) 13. P atil, S., Kim, D., P ark, S., Chai, Y.: Handwriting recognition in free space using WIMU-based hand motion analysis. Journal of Sensors 2016 (1), 3692876 (2016) 14. P ei, L., Ouyang, G.: Online recognition of handwritten characters from scalp- recorded brain activities during handwriting. Journal of Neural Engineering 18 (4), 046070 (2021) 15. Saini, M., Jain, A., Muthukrishnan, S.P ., Bhasin, S., Ro y , S., Kumar, L.: BiCur- Net: Premov ement EEG-based neural deco der for biceps curl tra jectory estimation. IEEE T ransactions on Instrumentation and Measurement 73 , 1–11 (2023) Title Suppressed Due to Excessive Length 11 16. Sc hirrmeister, R.T., Springen b erg, J.T., Fiederer, L.D.J., Glasstetter, M., Eggensp erger, K., T angermann, M., Hutter, F., Burgard, W., Ball, T.: Deep learn- ing with conv olutional neural netw orks for EEG deco ding and visualization. Human brain mapping 38 (11), 5391–5420 (2017) 17. Shiv ener, R., Carav ella, E.: W riting in Virtual Reality: Understanding Inv en- tion, Collaboration, and F riction in Hybrid Spaces. W ritten Comm unication p. 07410883251328315 (2025) 18. T ripathi, A., Gupta, A., Muthukrishnan, S.P ., Kumar, L., et al.: NeuroAiR: Deep learning framework for airwriting recognition from scalp-recorded neural signals. IEEE T ransactions on Instrumentation and Measurement 73 , 1–13 (2024) 19. T ripathi, A., Mondal, A.K., Kumar, L., AP , P .: SCLAiR: Supervised con trastive learning for user and device indep enden t airwriting recognition. IEEE Sensors Let- ters 6 (2), 1–4 (2021) 20. T ripathi, A., Mondal, A.K., Kumar, L., Prathosh, A.: ImAiR: Airwriting recogni- tion framew ork using image representation of IMU signals. IEEE Sensors Letters 6 (10), 1–4 (2022) 21. T ripathi, A., Prathosh, A., Muthukrishnan, S.P ., Kumar, L.: SurfMy oAiR: A sur- face electromy ograph y-based framew ork for airwriting recognition. IEEE T ransac- tions on Instrumentation and Measurement 72 , 1–12 (2023) 22. T ripathi, A., Prathosh, A., Muthukrishnan, S.P ., Kumar, L.: T ripCEAiR: A multi- loss minimization approach for surface EMG based airwriting recognition. Biomed- ical Signal Pro cessing and Control 85 , 104991 (2023) 23. V aidya, V., Prav anth, T., Viji, D.: Air writing recognition application for dyslexic p eople. In: 2022 International Mobile and Embedded T echnology Conference (MECON). pp. 553–558. IEEE (2022) 24. W u, M., Huang, K., Cai, Y., Hu, S., Zhao, Y., W ang, W.: Finger in camera sp eaks ev erything: Unconstrained air-writing for real-world. IEEE T ransactions on Cir- cuits and Systems for Video T echnology 34 (9), 8602–8613 (2024)