Paper deep dive
SwitchBraidNet: Quantisation-Aware Lightweight Architecture for Hybrid Brain-Computer Interface
Gourav Siddhad, Yogesh Kumar Meena
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 98%
Last extracted: 6/21/2026, 3:07:54 AM
Summary
The paper introduces SwitchBraidNet, a lightweight, quantization-aware deep learning architecture designed for hybrid Brain-Computer Interfaces (BCIs) combining Motor Imagery (MI) and Steady-State Visual Evoked Potentials (SSVEP). The model features a dual-path temporal braid, a squeeze-and-excitation spatial switch, and a log-variance readout layer. Evaluated on the OpenBMI dataset, SwitchBraidNet demonstrates superior robustness to low-bit quantization (INT8) compared to baselines like EEGNet and ShallowConvNet, achieving high MI accuracy (69.49% FP16) and SSVEP accuracy (93.48% FP32) with a minimal 3.03 KB footprint at INT8 precision.
Entities (9)
Relation Signals (5)
SwitchBraidNet â evaluatedon â OpenBMI
confidence 100% · through systematic quantisation-aware training on the OpenBMI dataset
SwitchBraidNet â uses â Dual-Path Temporal Braid
confidence 100% · The model employs a dual-path temporal braid to extract multiscale oscillatory features
SwitchBraidNet â uses â Spatial Switch
confidence 100% · The model employs... an adaptive squeeze-and-excitation spatial switch for electrode gating
SwitchBraidNet â uses â Log-Variance Readout
confidence 100% · a log-variance readout layer for direct band-power encoding
SwitchBraidNet â outperforms â ShallowConvNet
confidence 90% · SwitchBraidNet maintains the highest throughput and superior robustness at INT8 precision [compared to ShallowConvNet]
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Hybrid brain-computer interfaces (BCIs) that integrate motor imagery (MI) and steady-state visual evoked potentials (SSVEP) provide high-dimensional neural decoding but typically exceed the computational limits of embedded hardware. To address this, we propose SwitchBraidNet, a compact EEG classification architecture designed for low-power deployment. The model employs a dual-path temporal braid to extract multiscale oscillatory features, an adaptive squeeze-and-excitation spatial switch for electrode gating, and a log-variance readout layer for direct band-power encoding. Furthermore, through systematic quantisation-aware training on the OpenBMI dataset, we compared SwitchBraidNet against four established baselines across FP32, FP16, and INT8 precisions. Experimental results demonstrate superior efficiency and performance, achieving MI accuracy of 69.49% (FP16), SSVEP accuracy of 93.48% (FP32), and a hybrid information transfer rate of 64.82 bits/min (FP16). With an INT8 footprint of only 3.03 KB, SwitchBraidNet maintains high accuracy across varying numerical precisions, demonstrating its suitability for low-power embedded BCI deployment.
Tags
Links
- Source: https://arxiv.org/abs/2606.18816v1
- Canonical: https://arxiv.org/abs/2606.18816v1
Trouble viewing inline? Open PDF directly â
Full Text
34,110 characters extracted from source content.
Expand or collapse full text
SwitchBraidNet: Quantisation-Aware Lightweight Architecture for Hybrid Brain-Computer Interface Gourav Siddhad 1, and Yogesh Kumar Meena 1 1 Human-AI Interaction (HAIx) Lab, Indian Institute of Technology Gandhinagar, Gujarat, 382055, India Hybrid brainâcomputer interfaces (BCIs) that integrate motor imagery (MI) and steady-state visual evoked potentials (SSVEP) provide high-dimensional neural decoding but typically exceed the computational limits of embedded hardware. To address this, we propose SwitchBraidNet, a compact EEG classification architecture designed for low-power deployment. The model employs a dual-path temporal braid to extract multiscale os- cillatory features, an adaptive squeeze-and-excitation spatial switch for electrode gating, and a log-variance readout layer for direct band-power encoding. Furthermore, through sys- tematic quantisation-aware training on the OpenBMI dataset, we compared SwitchBraidNet against four established baselines across FP32, FP16, and INT8 precisions. Experimental re- sults demonstrate superior efficiency and performance, achiev- ing MI accuracy of 69.49% (FP16), SSVEP accuracy of 93.48% (FP32), and a hybrid information transfer rate of 64.82 bits/min (FP16). With an INT8 footprint of only 3.03 KB, SwitchBraid- Net maintains high accuracy across varying numerical preci- sions, demonstrating its suitability for low-power embedded BCI deployment. Deep Learning | EEG | Hybrid BCI | MI | Quantisation | SSVEP Correspondence: drgsiddhad@gmail.com 1: Introduction Brainâcomputer interfaces (BCIs) establish a direct com- munication pathway between the human nervous system and external devices, bypassing conventional neuromuscu- lar output channels (1). Among the spectrum of EEG-based BCI paradigms, motor imagery (MI) and steady-state visual evoked potentials (SSVEP) have emerged as the two most clinically and practically viable approaches. MI exploits the desynchronisation of ÎŒ-band (8â13 Hz) and ÎČ-band (13â 30 Hz) oscillations over the sensorimotor cortex during imag- ined limb movements (2). Whereas SSVEP leverages the phase-locked neural responses at occipital electrodes driven by flickering visual stimuli (3). Both paradigms have been validated in assistive technology applications ranging from wheelchair control to text spellers and robotic rehabilitation devices (4, 5). Their combination has been explored across a range of hybrid architectures, in- cluding gaze-MI systems that decouple visual target selection from motor-triggered confirmation (6), to further extend the scope of non-muscular communication. Nevertheless, both paradigms carry individual limitations: MI offers flexible voluntary control but suffers from low signal-to-noise ratio and high inter-subject variability, with BCI illiteracy rates as high as 53.7% reported on large-scale datasets (3), whereas SSVEP achieves high information throughput but demands continuous fixation on flickering stimuli, which is cognitively fatiguing for extended use (7). These complementary limitations have motivated the devel- opment of hybrid BCI (hBCI) systems that fuse signals from both paradigms to expand the effective command space and improve system robustness (4). The foundational argument for hybridisation rests on a key empirical observation: users who cannot effectively control an ERD-based BCI can of- ten operate an SSVEP-based one, and vice versa. This means that hybrid fusion substantially reduces the population of uni- versal BCI illiterates (3, 4). Recent work has extended this principle with peripheral visual field stimulation to improve user comfort while retaining SSVEP decoding accuracy (5), and with multi-modal combinations of MI, SSVEP, and overt spatial attention (OSA) to further enrich the neural feature space (8). Despite the rapidly growing literature on deep learning for EEG-based BCIs, three challenges have not been addressed jointly. First, the neural signatures of MI and SSVEP oc- cupy different spectral, spatial, and temporal subspaces, re- quiring feature representations that can accommodate hetero- geneous signal statistics within a single model (3, 7). Second, practical BCI deployment targets wearable or implantable hardware: microcontrollers, FPGAs, or neuromorphic chips, where memory, power, and compute budgets are severely constrained (9, 10). Importantly, while (10) simultaneously records MI and gaze signals in a live hybrid paradigm, (9) applies quantisation offline to a previously trained model, re- flecting a divide in the literature between real-time simulta- neous acquisition and post-hoc model compression. Third, quantising a trained model to low bit-widths (INT8 or below), as required for embedded deployment, introduces quantisa- tion noise that can disproportionately degrade architectures whose internal activations span a wide dynamic range (9). Existing benchmarks typically evaluate full-precision (FP32) models on a single paradigm, and the few works that con- sider model compression focus on post-training quantisa- tion (PTQ) rather than quantisation-aware training (QAT) (9), which has been shown to yield significantly better accuracy retention at INT8 and below (11). Also, hBCI evaluations rarely formalise the relationship between classification accu- Siddhad and Meena| arXiv | June 18, 2026 | 1â6 arXiv:2606.18816v1 [cs.HC] 17 Jun 2026 EEG Conv2d (1, 32) Conv2d (1, 64) deep fast Concat BN - ELU SE Block DW Conv Mixer 1x1 Conv LogVar FC Global Pool (Avg + Max) Linear + ReLU6 Linear + Hardsigmoid Skip SE Block Square Mean Log LogVar BN Temporal Braid Switch Dropout BN - ELU Fig. 1. Proposed SwitchBraidNet Architecture. The model employs a dual-path temporal braid using deep and fast convolutional kernels to extract multi-scale neural features. The spatial switch utilises a squeeze-and-excitation (SE) block to dynamically weigh electrode importance, followed by a hardware-efficient LogVarLayer that collapses the temporal dimension into stable, low-bit power features suitable for 8-bit QAT optimisation. racy and information transfer rate (ITR) across sequential vs simultaneous modes, leaving practitioners without analytical guidance on mode selection. The scoping literature on hBCIs confirms that simultaneous mode systems are the majority design choice, yet the ITR implications relative to sequential operation are not analytically derived or validated (12). To address these challenges, this work provides two major con- tributions, by: âą Proposing SwitchBraidNet, a new lightweight EEG deep learning architecture for the MI-SSVEP hybrid brain-computer interface (hBCI), designed for physio- logical and hardware constraints with end-to-end quan- tisation robustness. âą Presenting the first systematic quantisation-aware training benchmark on the OpenBMI dataset (3), com- paring five architectures at FP32, FP16, and INT8 across MI, SSVEP, and hybrid paradigms. 2: Materials and Methods This section describes the proposed model architecture, hBCI framework, quantisation strategy, dataset, preprocess- ing pipeline, experimental setup, and evaluation metrics used in this study. A. SwitchBraidNet. SwitchBraidNet is a lightweight archi- tecture proposed in this work, designed from the ground up with hardware-aware deployment in mind. It comprises three functional stages: a dual-path temporal braid, a spatial switch, and a log-variance readout head as shown in Fig 1. A.1. Dual-Path Temporal Braid. SwitchBraidNet employs two parallel temporal convolution paths operating at distinct temporal scales. A deep path uses a 1Ă 64 convolution (with padding 32) that captures slow oscillatory content (i.e., lower-frequency bands). A fast path uses a 1Ă 32 convolu- tion (with 16-pixel padding) tuned for faster transients. The outputs of both paths are concatenated along the channel di- mension to produce a joint representation, followed by batch normalisation (BN) and ELU activation. A.2. Spatial Switch via Squeeze-and-Excitation. The spatial stage applies a squeeze-and-excitation (SE) block (13) prior to spatial convolution. This serves as a learned switch that gates electrode contributions by their global relevance. The SE block computes both average-pooled and max-pooled statistics across the spatial-temporal dimensions, combines them, and passes the result through a two-layer bottleneck (reduction ratio r = 4) with ReLU6 and hardsigmoid acti- vations. The latter is advantageous for low-bit quantisation due to its piecewise-linear form. The SE output scales the feature maps via channel-wise multiplication before depth- wise spatial convolution (kernel CĂ 1, groups = 32). This reduces the spatial dimension from C to 1 while maintain- ing channel independence. Spatial BN, ELU activation, and 2-D dropout (0.2) follow. A 1Ă 1 pointwise mixer convolu- tion then projects from 32 to 16 channels with BN and ELU, acting as a learned feature recombination stage. A.3. Log-Variance Readout. SwitchBraidNet employs a Log- VarLayer that collapses the temporal dimension by comput- ing the mean squared amplitude across time-equivalent to a mean-power estimate, followed by a log transform with nu- merical stabilisation (Δ = 10 â6 ) and a BN layer: v = log(E t [x 2 ]+Δ),BN(v)(1) This design explicitly targets the power of neural oscillations, replicating the essential computation of logarithmic band- power features within a differentiable, quantisation-friendly layer. The BN within LogVarLayer stabilises the dynamic range of the output, which is critical for ensuring model ac- curacy when compressed to low-bit precision. The resulting 16-dimensional feature vector is passed to a linear classifica- tion head. 2 | arXivSiddhad and Meena| SwitchBraidNet: Quantisation-Aware Lightweight Architecture for Hybrid Brain-Computer Interface B Hybrid BCI Framework EEG EEG Search Select Visual AttentionMotor Imagery SSVEP (BCI) ERD (BCI) (a) EEG EEG Visual Attention Motor Imagery SSVEP (BCI) ERD (BCI) Search Select (b) Fig. 2. Framework of hybrid BCI (4): representing the data flow for (a) sequential and (b) simultaneous processing. B. Hybrid BCI Framework. A hBCI framework is im- plemented to overcome the inherent limitations of single- modality BCIs, namely the restricted command space of MI and the high cognitive fatigue or BCI-blindness with prolonged SSVEP use. A search-and-select paradigm de- signed to increase the number of discrete outputs is evalu- ated. Here, one modality serves as a high-level search/gating signal, while the other provides the final selection command. In the SSVEP-MI configuration, visual rhythmic entrainment (SSVEP) is used to navigate or search through interface lay- ers, while motor intent executes the final selection. In MI- SSVEP, asynchronous motor intent serves as a trigger to acti- vate the synchronous visual command space (SSVEP), effec- tively reducing false-positive activations during idle states. Two operational modes were considered for this framework. In sequential mode (Fig 2a), MI and SSVEP classifiers are applied to temporally distinct and non-overlapping trial win- dows. For e.g., a user first executes an MI trial and then attends to the SSVEP stimulus. In the simultaneous mode (Fig 2b), the user co-activates both neural sources within the same window. Since our architecture uses a shared backbone for both, the error patterns remain consistent across modes. However, the hybrid approach offers a significant advantage over single modalities by creating N MI Ă N SSV EP total command outputs. While the classification accuracy is math- ematically related, ITR vary significantly because simultane- ous execution is faster than sequential gating. C. Quantization-Aware Training. To evaluate the deploya- bility of each model under hardware resource constraints rep- resentative of embedded BCI devices, all five architectures were trained and evaluated across three numerical precisions: FP32, FP16, and INT8. Quantisation was implemented us- ing PyTorchâs native Torch Quantization framework with fake quantisation during the forward pass, enabling end-to- end QAT. Post-training quantisation (PTQ) was excluded be- cause compact models with narrow weight distributions are especially vulnerable to INT8 accuracy collapse without in- training adaptation (11). For FP32, no quantisation was ap- plied, whereas for FP16, native automatic mixed-precision was used and for INT8, the standard FBGEMM QAT config- uration was applied. This staged design allows each model to learn to compensate for quantisation noise during training, providing a controlled evaluation of accuracy degradation as a function of reduced numerical precision. SwitchBraidNetâsarchitecturalchoices(hardsigmoid, ReLU6, and batch-normalised log-power readout) were specifically selected to mitigate failure modes of low-bit QAT, such as activation collapse and weight clustering. The hardsigmoidâs bounded piecewise-linear form is empiri- cally favored for 8-bit quantisation (14), as it avoids the soft saturation of standard sigmoid that complicates scale estimation. D. Dataset and Preprocessing. All experiments used the publicly available OpenBMI dataset (3), comprising EEG recordings from 54 healthy subjects across two sessions, cov- ering a two-class MI task (left vs. right hand) and a four- class SSVEP task (5.45, 6.67, 8.57, and 12.00 Hz). From the original 62-channel, 1000 Hz recordings, paradigm-specific electrode subsets were selected (C=20 motor-cortex chan- nels for MI; 10 occipital channels for SSVEP) and down- sampled to 256 Hz to balance computational efficiency with signal integrity. MI signals were bandpass-filtered at 8â30 Hz to isolate ÎŒ/ÎČ rhythms, and SSVEP signals at 4â40 Hz to preserve stimulus-driven harmonics. Trials were segmented into 1 s windows with 0.5 s overlap, yielding input of shape (1ĂCĂ 256), providing sufficient frequency resolution for SSVEP decoding and temporal precision for MI-based event- related desynchronisation detection. E. Experimental Setup. The models were implemented us- ing PyTorch and MNE-Python, with training conducted over 50 epochs using the AdamW optimiser (LR=1eâ3). To ac- count for inter-session variability, predefined OpenBMI splits were utilised, ensuring calibration and evaluation trials re- mained strictly disjoint. For low-bitwidth simulations, QAT was facilitated by the Torch Quantisation library, which in- serted fake quantisation modules to model INT8 precision. All experiments were executed on an Nvidia RTX A2000 12GB GPU with CUDA 12.8, with model size estimation based on total parameter counts and their respective bit- widths. Five end-to-end deep learning architectures were evaluated, namely, DeepConvNet (15), EEGNet (16), Shal- lowConvNet (15), TSception (17) and the proposed Switch- BraidNet. F. Evaluation Metrics. Three primary metrics (Accuracy, Cohenâs Kappa, and ITR) were used to assess model perfor- mance across paradigms and quantisation levels. Classifica- tion accuracy was computed as the percentage of correctly classified trials over the held-out test set. To quantify classifi- cation performance while accounting for chance-level agree- ment, Cohenâs Kappa (Îș) was computed. To jointly quantify classification accuracy and command throughput, ITR (9, 10) was computed as follows. Let P denote classification accu- racy, N the number of target classes, and T the trial duration in seconds, then ITR is defined as (1) B = log 2 (N)+P log 2 (P)+(1âP)log 2 1âP Nâ1 (2) ITR = BĂ 60 T [bpm](3) Siddhad and Meena| SwitchBraidNet: Quantisation-Aware Lightweight Architecture for Hybrid Brain-Computer InterfacearXiv | 3 Where Eq. 2 gives the information content per trial in bits, and Eq. 3 converts it to a per-minute rate, giving bits per minute (bpm). Cases where P < 1/N are handled by clamp- ing B to zero to avoid undefined values. Additionally, the effective model size was estimated using the relation Size (KB) = [( P |parameters|Ă 4)/1024]Ă (b/32), where the factor b/32 scales the standard FP32 footprint by the quantised bit-width b â 32,16,8. Statistical signifi- cance was assessed using the Wilcoxon signed-rank test on paired observations across five models and three bit-width configurations (n = 15 pairs per comparison). The Friedman test was used for simultaneous comparison across quantisa- tion levels, followed by pairwise post hoc tests when signif- icant differences were found. All p-values were Bonferroni- corrected, with significance threshold of α=0.05. 3: Results This section presents the classification performance of five deep learning models (DeepConvNet, EEGNet, Shallow- ConvNet, TSception, and SwitchBraidNet) across three BCI paradigms (MI, SSVEP, and hBCI) and three numerical pre- cisions (FP32, FP16, and INT8). Single-modality and hBCI results are reported in Table 1, with model-parameter and memory statistics in Table 2. A. Single Modes. On the binary MI task, classification ac- curacy ranged from 63.79% to 69.49% across models and precision levels. SwitchBraidNet achieved the highest accu- racy at all three precisions, peaking at 69.49% under FP16, followed closely by EEGNet. Îș for SwitchBraidNet was 0.39 at both FP32 and FP16, indicating moderate agreement above chance. SwitchBraidNetâs ITR values represent a consistent improvement over EEGNet across all precisions. Shallow- ConvNet showed the largest drop under INT8, declining to 63.79% (â4.50% from FP32). On the four-class SSVEP task, all models achieved substantially higher accuracy than on MI. EEGNet attained the highest metrics across all preci- sions. SwitchBraidNet performed comparably to EEGNet, with a negligible gap of 0.17%. DeepConvNet also per- formed at a level comparable to SwitchBraidNet. Shallow- ConvNet achieved the lowest SSVEP accuracy among the five models, with further degradation observed at INT8. B. Hybrid Modes. Sequential and simultaneous hBCI modes produced identical classification accuracy, F1-score, and Îș across all models and precision levels (p = 1.0 for all three metrics), confirming that the two modes share the same decision quality. The ITR differed substantially be- tween modes: the simultaneous mode achieved a signifi- cantly higher ITR (59.0 vs. 29.5 bpm, p< 0.001), as co- acquiring signals within an overlapping window halves the effective trial duration from 2 s to 1 s (see Eq. 2). Rela- tive to the MI-alone baseline, the hybrid system achieved a statistically significant improvement in ITR (+23.8 bpm, p< 0.001). Relative to the SSVEP-alone baseline, all hybrid metrics declined significantly (ITR:â57.9 bpm, p< 0.001). 32168 Bit Depth (Bits) 30 32 34 36 38 40 42 44 46 ITR (bits/min) DeepConvNet EEGNet ShallowConvNet TSception SwitchBraidNet Fig. 3. ITR vs. bit-depth across all architectures. SwitchBraidNet main- tains the highest throughput and superior robustness at INT8 precision. C. Quantization Robustness. Quantisation level had a statistically significant effect on MI-based paradigms (Fried- man test, p< 0.05 for Acc, F1, ITR, and Îș), with perfor- mance declining monotonically from 32-bit to 8-bit. No individual bit-width pair reached significance after Bonfer- roni correction. SSVEP-only models were robust to quan- tisation (p = 0.074). The FP32-to-FP16 transition was ef- fectively lossless for all models except ShallowConvNet. SwitchBraidNet gained 0.16% on MI and lost 0.06% on SSVEP. Under INT8, SwitchBraidNet sustained the smallest performance degradation across both paradigms (0.21% on MI, 0.08% on SSVEP),â20Ă smaller than ShallowConvNet (4.50% and 3.75%, respectively). SwitchBraidNet achieved the best MI accuracy at INT8 (69.12%) with a memory foot- print of 3.03 KB; EEGNet achieved the best SSVEP accuracy at INT8 (93.67%) at 2.24 KB. As shown in Fig. 3, all architectures remain stable from FP32 to FP16 but diverge at INT8. ShallowConvNetâs ITR drops from 37 to 30 bpm, while SwitchBraidNet maintains the highest baseline ITR (â46 bpm) with negligible degradation. EEGNet and TSception also exhibit high robustness to quan- tisation, though at lower absolute ITR levels than Switch- BraidNet. 4: Discussion Across all experimental conditions, two trends emerge from the results. First, SwitchBraidNet achieves the highest ac- curacy on MI and both hybrid hBCI configurations across all precisions, and attains accuracy and ITR that are statis- tically comparable to EEGNet on SSVEP, despite a similarly compact memory footprint (3.03 KB vs 2.24 KB at INT8). Second, ShallowConvNet exhibits the strongest sensitivity to quantisation, whereas SwitchBraidNet demonstrates superior robustness, degrading by less than 0.25% across all tested conditions. A. Single-Mode Performance. The narrow inter-model accuracy gap on MI (<6%) reflects the inherently low spa- tial resolution of motor-imagery signals. SwitchBraidNetâs marginal advantage stems from its dual-path temporal braid, 4 | arXivSiddhad and Meena| SwitchBraidNet: Quantisation-Aware Lightweight Architecture for Hybrid Brain-Computer Interface B Hybrid vs. Single-Modality Trade-offs Table 1. Performance metrics for single modality paradigms and hybrid BCI systems in sequential and simultaneous modes DeepConvNetEEGNetShallowConvNetTSceptionSwitchBraidNet Paradigm3216832168321683216832168 MI Acc68.57 68.75 68.39 69.08 68.67 68.04 68.29 68.58 63.79 65.73 66.09 65.48 69.33 69.49 69.12 F167.91 68.44 68.35 69.06 68.46 68.02 68.08 68.58 62.50 65.65 66.05 65.47 69.23 69.27 68.92 ITR6.126.245.996.466.185.765.936.123.334.364.564.226.646.756.50 Kappa0.370.380.370.380.370.360.370.370.280.310.320.310.390.390.38 SSVEP Acc93.32 93.26 93.24 93.65 93.71 93.67 86.77 86.33 83.02 92.21 91.99 91.66 93.48 93.42 93.40 F193.33 93.29 93.25 93.67 93.73 93.69 86.78 86.34 82.80 92.24 92.01 91.66 93.49 93.44 93.41 ITR92.42 92.21 92.15 93.49 93.70 93.57 73.58 72.46 64.42 88.90 88.22 87.22 92.94 92.74 92.69 Kappa0.910.910.910.920.920.920.820.820.770.900.890.890.910.910.91 Acc63.99 64.12 63.77 64.69 64.35 63.73 59.25 59.20 52.96 60.61 60.79 60.01 64.81 64.92 64.56 HybridF163.38 63.84 63.74 64.68 64.17 63.73 59.08 59.21 51.76 60.56 60.77 60.01 64.73 64.73 64.38 MI-SSVEP / Kappa0.590.590.590.600.590.590.530.530.460.550.550.540.600.600.60 SSVEP-MIITR (Seq) 31.39 31.53 31.14 32.16 31.79 31.11 26.42 26.38 20.46 27.81 28.00 27.20 32.29 32.41 32.02 ITR (Sim) 62.78 63.05 62.29 64.32 63.58 62.22 52.85 52.75 40.91 55.62 55.99 54.40 64.58 64.82 64.04 Table 2. Comparison of model complexity Memory (KB) ModelParameters32 Bit16 Bit8 Bit DeepConvNet87,552 342.00 171.00 85.50 EEGNet2,2928.954.482.24 ShallowConvNet17,96270.1635.08 17.54 TSception81,617 318.82 159.41 79.70 SwitchBraidNet3,10612.136.073.03 which simultaneously captures ÎŒ/ÎČ-band power envelopes across parallel branches without requiring explicit filter-bank segmentation. The LogVarLayerâs mean-power readout fur- ther specialises the network towards band-power features, which are the physiologically motivated correlates of MI. ShallowConvNet also targets band-power but underperforms under INT8 quantisation, suggesting that its pooling window is more sensitive to precision-induced weight perturbations. The substantially higher SSVEP accuracies across all models reflect the high SNR of frequency-tagged responses. EEG- Netâs separable convolutional architecture compactly en- codes the spectral-spatial structure of steady-state signals, which explains its lead on this task. SwitchBraidNet per- formed almost identically to EEGNet on SSVEP (0.17% gap), indicating that its architecture generalises well be- yond its primary MI design target. ShallowConvNetâs lower SSVEP performance and INT8 sensitivity further suggest that its spatial pooling window lacks the granularity needed for multi-frequency discrimination. B. Hybrid vs. Single-Modality Trade-offs. The hybrid paradigm introduces an accuracy cost relative to single- modality SSVEP while expanding the command space. Un- der FP32, the best observed SSVEP accuracy (93.65%) far exceeds the best hybrid accuracy (64.81%), which is ex- pected: fusing the harder MI task (â69%) with SSVEP across a higher-dimensional joint decision boundary in- evitably increases the error rate. Nevertheless, the hybrid paradigm offers practical advan- tages for real-world deployment. First, combining two-class MI with four-class SSVEP yields eight commands, enabling richer control without additional hardware. Second, periodi- cally offloading to the MI channel reduces ocular strain from sustained SSVEP fixation (18, 19). Third, grounding part of the command vocabulary in SSVEP improves reliability against MI non-stationarity (20). Fourth, the dual-modality design provides a functional fallback for users whose MI or SSVEP signal quality is situationally impaired. The simultaneous hybrid ITR falls short of single-modality SSVEP because the accuracy reduction dominates the command-space expansion in the ITR formulation.For practical deployment, the sequential mode may be preferred when simultaneous co-activation is cognitively prohibitive, whereas the simultaneous mode remains attractive where command bandwidth and user comfort across extended ses- sions are the primary design objectives. C. Quantization Robustness. The lossless FP32-to-FP16 transition observed for most models is consistent with the known regularising effect of reduced-precision arithmetic, which introduces small stochastic perturbations during train- ing analogous to mild noise injection (11). This is consistent with the marginal accuracy gain of SwitchBraidNet at FP16 on MI (0.16%). Under INT8 - the most practically relevant precision for microcontroller and FPGA deployment (e.g., ARM Cortex- M) - SwitchBraidNetâs â20Ă smaller degradation relative to ShallowConvNet indicates that its temporal pathways and spatial switching preserve the features required for MI and SSVEP decoding under constrained numerical preci- sion.SSVEP modelsâ overall robustness to quantisation (p = 0.074) further suggests that frequency-domain features are less sensitive to weight precision than the spatial features underpinning MI decoding. For SSVEP-only deployment, EEGNet (2.24 KB at INT8) represents the optimal choice by accuracy. For hybrid de- ployment, SwitchBraidNetâs consistent top-2 performance Siddhad and Meena| SwitchBraidNet: Quantisation-Aware Lightweight Architecture for Hybrid Brain-Computer InterfacearXiv | 5 246810121416 #Targets (SSVEP) 65 70 75 80 85 90 Accuracy (%) for BCI #Targets (MI)=2 #Targets (MI)=3 #Targets (MI)=4 Fig. 4. Minimum MI classification accuracy required for the hBCI to ex- ceed the ITR of a pure SSVEP system, as a function of the number of SSVEP targets. SSVEP accuracy is assumed to be perfect. across both modalities and its small footprint make it the most versatile model. D. ITR Threshold Analysis for hBCI. Fig. 4 maps the min- imum MI accuracy required for the hBCI ITR to exceed that of a pure SSVEP system, for N MI â2,3,4 and assuming perfect SSVEP accuracy as a theoretical upper bound. For N MI = 2, the required minimum at N SSVEP = 2 is 81.1%, rising and asymptoting as N SSVEP increases well above the â69% MI accuracy observed across all five architectures. This finding is consistent with OâDoherty et al. (6), who reported an analogous 86% threshold for a gaze-MI hy- brid. Moreover, this bound is strict; at empirically observed SSVEP accuracies (â93.5%), the SSVEP baseline ITR de- creases, further narrowing the gap. Hybrid deployment is therefore justified primarily by complementary advantages: expanded command vocabulary, reduced visual fatigue, and fallback resilience, rather than by net ITR gain over SSVEP alone. 5: Conclusion This paper presented a systematic evaluation of five deep learning architectures for hybrid EEG-based BCIs under QAT, utilising the OpenBMI dataset across MI, SSVEP, and hBCI paradigms. The central contribution is SwitchBraid- Net, a lightweight three-stage architecture comprising a dual- path temporal braid, a squeeze-and-excitation spatial switch, and a log-variance layer. It achieves the highest MI and hBCI classification accuracy across all tested precisions (INT8 ac- curacy: 69.12% MI, 64.81% hBCI; footprint: 3.03 KB), with negligible INT8 degradation (0.21%) and ITRs of 32.41 and 64.82 bpm in sequential and simultaneous modes, respec- tively. On SSVEP, EEGNet leads marginally, but Switch- BraidNet remains statistically competitive, with both exhibit- ing negligible INT8 degradation. The QAT benchmark and ITR framework presented here provide a reproducible foun- dation for future research in neural signal processing and hardware-aware deep learning, with a natural extension be- ing the incorporation of a third modality, such as P300, to further expand the command space. Bibliography 1. Jonathan R Wolpaw, Niels Birbaumer, William J Heetderks, Dennis J McFar- land, P Hunter Peckham, Gerwin Schalk, Emanuel Donchin, Louis A Qua- trano, Charles J Robinson, Theresa M Vaughan, et al. Brain-computer inter- face technology: a review of the first international meeting. IEEE Transactions on Rehabilitation Engineering, 2000. doi: 10.1109/tre.2000.847807. 2. Gert Pfurtscheller and Christa Neuper.Motor imagery and direct brain- computer communication. Proceedings of the IEEE, 2001. doi: 10.1109/5. 939829. 3. Min-Ho Lee, O-Yeon Kwon, Yong-Jeong Kim, Hong-Kyung Kim, Young-Eun Lee, John Williamson, Siamac Fazli, and Seong-Whan Lee. Eeg dataset and openbmi toolbox for three bci paradigms: an investigation into bci illiteracy. GigaScience, 2019. doi: 10.1093/gigascience/giz002. 4. Gert Pfurtscheller. The hybrid bci. Frontiers in Neuroscience, 2010. doi: 10.3389/fnpro.2010.00003. 5. Ruoqing Zhang, Zhaohui Li, Xiaohu Pan, Hongyan Cui, and Xiaogang Chen. Hybrid bci for upper limb rehabilitation: integrating mi with peripheral field ssvep stimulation. Journal of Neuroscience Methods, 2025. doi: 10.1016/j. jneumeth.2025.110537. 6. Darren OâDoherty, Yogesh Kumar Meena, Haider Raza, Hubert Cecotti, and Girijesh Prasad. Exploring gaze-motor imagery hybrid brain-computer in- terface design. 2014 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2014. doi: 10.1109/bibm.2014.6999180. 7. Wenwei Luo, Wanguang Yin, Quanying Liu, and Youzhi Qu. A hybrid brain- computer interface using motor imagery and ssvep based on convolutional neural network. Brain-Apparatus Communication: A Journal of Bacomics, 2023. doi: 10.1080/27706710.2023.2258938. 8. Chao Zhang, Guojing Li, Xiaopei Wu, and Xiangping Gao. A novel hybrid brainâcomputer interface integrating motor imagery and multiple visual stimuli. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2025. doi: 10.1109/tnsre.2025.3591616. 9. Hubert Cecotti, Dalvir Dhaliwal, Hardip Singh, and Yohesh Kumar Meena. Post-training quantization in brain-computer interfaces based on event-related potential detection. 2024 IEEE International Conference on Systems, Man, and Cybernetics (SMC), 2024. doi: 10.1109/smc54092.2024.10831138. 10. Yogesh Kumar Meena, Hubert Cecotti, KongFatt Wong-Lin, and Girijesh Prasad. Towards increasing the number of commands in a hybrid brain- computer interface with combination of gaze and motor imagery. 2015 37th Annual International Conference of the IEEE Engineering in Medicine and Bi- ology Society (EMBC), 2015. doi: 10.1109/embc.2015.7318410. 11. Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. doi: 10.1109/cvpr.2018.00286. 12. Matheus G. Mussi and Kim D. Adams. Eeg hybrid brain-computer interfaces: A scoping review applying an existing hybrid-bci taxonomy and considerations for pediatric applications. Frontiers in Human Neuroscience, 2022. doi: 10. 3389/fnhum.2022.1007136. 13. Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. doi: 10.1109/cvpr.2018.00745. 14. Yash Bhalgat, Jinwon Lee, Markus Nagel, Tijmen Blankevoort, and Nojun Kwak. Lsq+: Improving low-bit quantization through learnable offsets and better initialization. 2020 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition Workshops (CVPRW), 2020. doi: 10.1109/cvprw50498.2020. 00356. 15. Robin Tibor Schirrmeister, Jost Tobias Springenberg, Lukas Dominique Josef Fiederer, Martin Glasstetter, Katharina Eggensperger, Michael Tangermann, Frank Hutter, Wolfram Burgard, and Tonio Ball. Deep learning with convo- lutional neural networks for eeg decoding and visualization. Human Brain Mapping, 2017. doi: 10.1002/hbm.23730. 16. Vernon J Lawhern, Amelia J Solon, Nicholas R Waytowich, Stephen M Gor- don, Chou P Hung, and Brent J Lance. Eegnet: a compact convolutional neural network for eeg-based brainâcomputer interfaces. Journal of Neural Engineering, 2018. doi: 10.1088/1741-2552/aace8c. 17. Yi Ding, Neethu Robinson, Su Zhang, Qiuhao Zeng, and Cuntai Guan. Ts- ception: Capturing temporal dynamics and spatial asymmetry from eeg for emotion recognition. IEEE Transactions on Affective Computing, 2022. doi: 10.1109/taffc.2022.3169001. 18. Dimitra Makri, Christina Farmaki, and Vangelis Sakkalis. Visual fatigue effects on steady state visual evoked potential-based brain computer interfaces. 2015 7th International IEEE/EMBS Conference on Neural Engineering (NER), 2015. doi: 10.1109/ner.2015.7146562. 19. Pablo DĂez, Lorena Orosco, Agustina GarcĂ©s Correa, and Luciano Carmona. Assessment of visual fatigue in ssvep-based brain-computer interface: a com- prehensive study. Medical & Biological Engineering & Computing, 2024. doi: 10.1007/s11517-023-03000-z. 20. Pradeep Shenoy, Matthias Krauledat, Benjamin Blankertz, Rajesh P N Rao, and Klaus-Robert MĂŒller. Towards adaptive classification for bci. Journal of Neural Engineering, 2006. doi: 10.1088/1741-2560/3/1/r02. 6 | arXivSiddhad and Meena| SwitchBraidNet: Quantisation-Aware Lightweight Architecture for Hybrid Brain-Computer Interface