BioTrain: On-Device Fine-Tuning for Biosignals

Updated 4 July 2026

BioTrain is a compiler-assisted on-device fine-tuning framework for biosignal AI that enables full-network training under sub-megabyte memory and sub-50 mW power constraints.
It employs a specialized compiler stack to transform full training graphs into efficient, bare-metal C code optimized for GAP9-class hardware, ensuring privacy and reliable personalization.
The framework achieves an 8.1x memory reduction and outperforms shallow adaptations, demonstrating robust Day-1 and longitudinal improvements in EEG and EOG models.

BioTrain is a compiler-assisted on-device fine-tuning framework for biosignal AI that enables full-network fine-tuning of state-of-the-art biosignal models on MCU-class edge devices under sub-megabyte peak memory and sub-50 mW power constraints (Wang et al., 14 Apr 2026). It addresses post-deployment degradation caused by cross-subject and cross-session variability in biosignals, and it does so by compiling full training graphs—forward pass, backward pass, and weight updates—into bare-metal C for GAP9-class hardware. In the reported implementation, BioTrain extends Deeploy with optimized gradient kernels from PULPTrainLib and demonstrates fully on-chip personalization for EEG and EOG models, including Day-1 calibration and longitudinal adaptation to signal drift (Wang et al., 14 Apr 2026).

1. Problem setting and adaptation objective

BioTrain is motivated by two forms of domain shift in wearable biosignal sensing: cross-subject shift, where signals differ across individuals because of anatomy, electrode placement, skin impedance, and behavioral differences, and cross-session shift, where signals drift over time because of perspiration, sensor displacement, changing impedance, and day-to-day variation (Wang et al., 14 Apr 2026). In that setting, an offline model trained on pooled subjects can perform poorly when deployed to a new user or when used over long time horizons.

The framework targets on-device adaptation for two reasons stated explicitly in the paper: preserving user privacy and ensuring system reliability (Wang et al., 14 Apr 2026). It also rejects the assumption that shallow adaptation is sufficient. Prior on-device methods typically update only the last classification layer, a sparse subset of weights, or shallow/self-supervised statistics, but the reported results indicate that these strategies often fail to recover the accuracy of full fine-tuning when domain shift is large, especially for EEG (Wang et al., 14 Apr 2026).

Two deployment scenarios structure the methodology. In Day-1 calibration, a model is pre-trained offline using leave-one-subject-out on $N-1$ subjects, then fine-tuned for a new unseen subject on the first session’s training portion (80%) and evaluated on the remaining 20%. In longitudinal adaptation, the model is updated session by session on later sessions $S_2, S_3, \dots$ , again fine-tuning on the first 80% and testing on the remaining 20% (Wang et al., 14 Apr 2026). This makes BioTrain a personalization framework rather than a generic embedded training library.

A common misconception is that BioTrain is merely a reduced form of last-layer calibration. The paper’s central claim is stronger: full-network fine-tuning matters for both Day-1 new-subject calibration and longitudinal adaptation, and the main obstacle is not the desirability of full backpropagation but its memory footprint on microcontrollers (Wang et al., 14 Apr 2026).

2. Compiler stack and execution model

BioTrain is not presented as a standalone optimization rule; it is a deployment compiler stack for end-to-end training on a microcontroller (Wang et al., 14 Apr 2026). The system is organized into a frontend, midend, and backend.

The frontend supports CNN backpropagation operators, including normalization, replaces unsupported training-graph components with synthesized gradient subgraphs, decomposes gradients into separate input-gradient and weight-gradient paths, and decomposes normalization into statistic computation and normalization computation (Wang et al., 14 Apr 2026). This stage turns a training graph into a form that can be scheduled under scratchpad constraints.

The midend performs tiling and static memory allocation over the full training graph. It uses OR-Tools to solve a constrained optimization problem for scratchpad-aware scheduling, focuses on temporal tiling for long 1D biosignal sequences with few channels, orchestrates DMA-compute overlap, and minimizes off-chip transfers (Wang et al., 14 Apr 2026). The emphasis on temporal tiling is a direct consequence of the target workload: EEG and EOG inputs are long sequences rather than large 2D tensors.

The backend generates optimized kernels for both forward and backward computation. It reuses Deeploy and PULP-NN kernels for the forward path, adds gradient kernels adapted from PULPTrainLib, including convolution, normalization, pooling, and activation gradients, and modifies those kernels to operate under the compiler’s tiling scheme (Wang et al., 14 Apr 2026). The result is an executable training pipeline rather than a hand-written runtime.

This organization has an important methodological implication. BioTrain treats on-device learning as a compilation and memory-orchestration problem, not only as a model-design problem. The paper therefore positions compiler support, kernel specialization, and scheduling as first-class components of personalization.

3. Memory reduction mechanisms and normalization strategy

BioTrain identifies two primary memory bottlenecks in MCU training: activation storage for backpropagation and batch statistics / cross-sample dependencies (Wang et al., 14 Apr 2026). Its memory-efficiency strategy combines gradient accumulation, Group Normalization, temporal tiling, Welford aggregation, and static memory allocation.

Gradient accumulation allows the system to run multiple single-sample forward/backward passes and accumulate gradients until reaching an effective batch size (Wang et al., 14 Apr 2026). This preserves the optimization behavior of a larger batch while keeping peak activation memory low. The mechanism is essential because a conventional mini-batch would exceed available SRAM.

A central architectural substitution is GroupNorm instead of BatchNorm. The paper argues that Batch Normalization is problematic for edge training with gradient accumulation because it depends on cross-sample batch statistics, which creates synchronization and memory complications when processing small chunks or even single samples (Wang et al., 14 Apr 2026). Group Normalization depends only on per-sample statistics, works better with small effective batch sizes, avoids cross-sample coupling, and preserves stable optimization in memory-constrained training.

Temporal tiling exploits the structure of biosignal inputs. For convolution weight gradients, if $T_k$ is the temporal tile at step $k$ , BioTrain computes partial gradients as

$\nabla W^{(k)} = \mathrm{im2col}(X_{T_k})^\top \cdot dY_{T_k},$

with each tile contributing to a persistent gradient buffer (Wang et al., 14 Apr 2026). For input gradients $dX$ , the corresponding $dY$ tiles are extended with a kernel-dependent halo so that resulting $dX$ tiles are non-overlapping and schedulable.

For normalization layers, the framework uses a cross-tile Welford aggregation scheme to compute mean and variance from tile-local statistics without storing the whole sequence in scratchpad (Wang et al., 14 Apr 2026). This is a direct response to the fact that normalization otherwise introduces long-range dependencies over the temporal dimension.

The aggregate effect is substantial. For fully on-chip backpropagation on GAP9, BioTrain reduces the memory footprint from 5.4 MB to 0.67 MB for a conventional full-network fine-tuning setup using Batch Normalization with batch size 8, an 8.1x reduction (Wang et al., 14 Apr 2026). This is the enabling condition for fitting full-network training into a device with 1.5 MB L2 SRAM.

4. Hardware target, datasets, and evaluation protocol

BioTrain is evaluated on the GAP9 MCU, which includes a 9-core RISC-V compute cluster plus controller core, 128 kB L1 SRAM, 1.5 MB L2 SRAM, and DMA-based transfers between L2 and L1 (Wang et al., 14 Apr 2026). Power is measured with the Nordic Power Profiler Kit II.

The paper studies two biosignal datasets. The EEG dataset comprises 5 subjects, 4 sessions per subject, 8 EEG channels, and binary classification of tongue vs. rest, with 3.8 s trials and $T=1900$ ; the backbone is MI-BMINet with 7.9k parameters (Wang et al., 14 Apr 2026). The EOG dataset comprises 5 subjects, 2 sessions per subject, 3 dry electrodes derived into horizontal and vertical EOG channels, and 11-class eye-movement classification, with 2.0 s trials and $T=1000$ ; the backbone is EpiDeNet with 4.1k parameters (Wang et al., 14 Apr 2026).

Four adaptation strategies are compared:

No-FT: zero-shot, no fine-tuning
LP: last-layer probing/update only
Full-FT: full-network backpropagation with the original architecture
Edge-FT: BioTrain’s edge-optimized full fine-tuning, using GN and compiler-driven memory orchestration (Wang et al., 14 Apr 2026)

This evaluation protocol is designed to separate algorithmic adaptation quality from deployability. Full-FT serves as the unconstrained reference, LP as the shallow-adaptation baseline, and Edge-FT as the resource-constrained realization of full-network personalization.

5. Accuracy, throughput, and deployment characteristics

The reported results show that full fine-tuning is materially stronger than No-FT and typically stronger than LP, while BioTrain’s Edge-FT matches or exceeds standard Full-FT under the evaluated conditions (Wang et al., 14 Apr 2026).

Setting	Strategy	Accuracy
EEG Day-1	No-FT	50.8 ± 8.8
EEG Day-1	LP	74.8 ± 11.4
EEG Day-1	Full-FT	80.0 ± 11.4
EEG Day-1	Edge-FT	86.4 ± 7.9
EEG longitudinal/session-wise	No-FT	49.8 ± 3.8
EEG longitudinal/session-wise	LP	76.2 ± 7.6
EEG longitudinal/session-wise	Full-FT	78.7 ± 7.3
EEG longitudinal/session-wise	Edge-FT	83.9 ± 13.3
EOG Day-1	No-FT	78.1 ± 18.1
EOG Day-1	LP	83.3 ± 13.0
EOG Day-1	Full-FT	88.7 ± 10.3
EOG Day-1	Edge-FT	87.7 ± 10.4
EOG longitudinal/session-wise	No-FT	84.1 ± 11.2
EOG longitudinal/session-wise	LP	86.5 ± 10.7
EOG longitudinal/session-wise	Full-FT	89.1 ± 8.8
EOG longitudinal/session-wise	Edge-FT	87.2 ± 7.6

On EEG, Edge-FT improves over No-FT by 35.6 percentage points in the Day-1 setting and exceeds LP by about 11.6 points (Wang et al., 14 Apr 2026). On EOG, gains are smaller because the task is easier and baseline performance is already higher, but Edge-FT remains close to Full-FT (Wang et al., 14 Apr 2026). The paper further states that LP can show unstable convergence or adaptation delay across sessions, whereas Edge-FT remains more stable in EEG longitudinal adaptation.

The systems results are equally central because BioTrain is defined by feasibility on hardware rather than offline accuracy alone.

Model / method	Peak memory	Key deployment metrics
EEG LP	0.19 MB	360.0 ms, 45.2 mW, 16.08 mJ, 0.38 GFLOPs/s
EEG Full-FT	5.36 MB	not feasible on GAP9 L2 as-is
EEG Edge-FT	0.67 MB	469.6 ms, 43.8 mW, 20.16 mJ, 0.89 GFLOPs/s
EOG LP	0.11 MB	34.4 ms, 50.4 mW, 1.76 mJ, 0.21 GFLOPs/s
EOG Full-FT	2.24 MB	not feasible on GAP9 L2 as-is
EOG Edge-FT	0.28 MB	93.6 ms, 48.2 mW, 4.48 mJ, 0.22 GFLOPs/s

On GAP9, BioTrain enables efficient on-device training throughput of 17 samples/s for EEG and 85 samples/s for EOG within a power envelope below 50 mW (Wang et al., 14 Apr 2026). The framework therefore converts full-network fine-tuning from an offline-only capability into an on-chip adaptation workflow. The paper also estimates that a 320 mAh, 3.7 V battery can support about 211 EEG training sessions and about 951 EOG training sessions, assuming 40 epochs over 200 samples (Wang et al., 14 Apr 2026).

Several limitations bound these claims. The implementation is focused on CNN-based biosignal models, training is described in FP32, and evaluation covers only EEG and EOG rather than other biosignals such as EMG or multimodal wearable sensing (Wang et al., 14 Apr 2026). Generalization beyond the compiler-supported topology and tiling assumptions is therefore not established.

The term BioTrain is not unique across the literature. In the supplied corpus, the closest direct naming conflict is BFTrainer, a system for low-cost training of neural networks on unfillable supercomputer nodes, which is also described as BioTrain in the query context (Liu et al., 2021). BFTrainer addresses FCFS scheduling inefficiency on systems such as Summit by allocating transient idle nodes to fully preemptable elastic DNN Trainers and formalizing the rescaling problem as a mixed-integer linear program solved with Gurobi (Liu et al., 2021). Its subject matter is high-performance computing resource allocation rather than biosignal personalization.

Other papers exhibit a related “train-efficiently” theme without introducing a framework named BioTrain. In terahertz communications, "Energy Efficient Beamforming Training in Terahertz Communication Systems" proposes EETBF, which decomposes optimization into EETBF-BT for history-guided beam selection and EETBF-PA for Q-learning-based training power assignment, achieving the lowest training latency and power consumption among the compared schemes (Shen et al., 2024). In quantitative MRI, "Only-Train-Once MR Fingerprinting for Magnetization Transfer Contrast Quantification" introduces OTOM, a bidirectional-LSTM framework that decodes MTC-MRF fingerprints across variable-length and variable-pattern acquisition schedules without schedule-specific retraining (Kang et al., 2022). In XR training research, "XR: Enabling training mode in the human brain" treats training effectiveness as a function of the right vector of sensory cues and the right pedagogical structure rather than maximal realism (Lépinard et al., 2019).

These usages are conceptually adjacent only at a high level. BioTrain in the strict sense denotes the 2026 biosignal framework for compiler-driven, full-network on-device fine-tuning under sub-MB and sub-50 mW constraints (Wang et al., 14 Apr 2026). BFTrainer, EETBF, OTOM, and the XR training-mode formulation address distinct problems—supercomputer backfill holes, terahertz beam alignment, schedule-agnostic MR fingerprint decoding, and XR pedagogy, respectively—and should not be conflated with the biosignal system despite the lexical overlap or “train once / training efficiency” resonance across the broader literature.