A tiny parameter count is only useful if the full pipeline fits. LiMO combines configurable quantization and layer fusion to run ECG classification entirely on an nRF52840 MCU.
#TinyML#ECG
https://t.co/yN36uMjjmH
EEG channels have a spatial layout. Our preprint preserves that layout in 2D maps before learning cross-subject features with a ResNet-BiLSTM backbone. Evaluated on MODMA.
#EEG#TransferLearning
https://t.co/NSg6eDWeJH
How a memory cell is read shapes the computation. Our FeRAM work studies charge-mode readout in capacitive crossbars, with ternary weights and speech inference simulated from measured device behavior.
#FeRAM#ComputeInMemory
https://t.co/orZ80yXt9d
A defense that fits in software may still be expensive on a device. Our FPGA ECG study evaluates adversarial robustness alongside quantization and inference cost.
#ECG#EdgeAI
https://t.co/akkvZumAfJ
What if image analysis helped choose what a camera sends next? FICE feeds lesion detection and localization back into selective high-resolution imaging, evaluated on public endoscopy datasets.
#Endoscopy#EdgeAI
https://t.co/s1OceaTiLR
In a hybrid pathology image, which pixels remain original? Our telepathology manuscript preserves selected native-resolution tiles while reconstructing the background. Evaluated on CAMELYON17.
#DigitalPathology#MedicalImaging
https://t.co/CWbYY6NCDR
Sending every sensor sample has an energy cost. Our optical-capsule preprint compares on-device bleeding-flow classification with raw-data transmission, using controlled in-vitro gastric experiments.
#EdgeAI#Biosensors
https://t.co/zWh0QaNte9
Switching radio protocols takes time. Our Enhanced-BLE preprint measures BLE-to-ESB handover and restoration of BLE operation on the nRF54L15, alongside throughput and energy.
#IoT#WirelessCommunication
https://t.co/LUkmvWGQPw
Denoising in two stages: our heart-sound manuscript uses a reference channel to cancel correlated noise, then wavelet denoising for residual noise while preserving S1/S2 transients.
#HeartSounds#SignalProcessing
https://t.co/CWGruLeHAQ
Compression has a memory cost too. Our CapsuleCoder manuscript uses a streaming design that avoids buffering full feature maps while allocating bits according to lesion relevance.
#Endoscopy#EdgeAI
https://t.co/TVTYV4jlrz
ECG and heart sounds capture different parts of a heartbeat. Our HAM-CFN preprint uses bidirectional cross-modal attention to exchange multiscale features between the two signals.
#ECG#MultimodalAI
https://t.co/wcRD1V446N
The same MRI can describe individual regions and relationships between them. Our temporal lobe epilepsy study combines regional morphology with personalized brain-network features for classification.
#Neuroimaging#BrainNetworks
https://t.co/UM7wO5hd1S
A sigmoid is also a cumulative distribution function. Our hardware approximator uses that connection to build a probabilistic implementation, with indirect quantization and a latency-optimized design.
#FPGA#EfficientAI
https://t.co/gbJ3D9uM3I
For multimodal MRI, the evaluation setting matters. Our autism classification study fuses structural and functional features, tested with five-fold cross-validation on the ABIDE NYU dataset.
#Neuroimaging#MultimodalAI
https://t.co/q1NJpS3vXY
A small speech model starts before the network. Our FPGA PoolFormer design pairs spiking feature extraction with a compact classifier, evaluated on Google Speech Commands.
#TinyML#SpeechRecognition
https://t.co/wLAuymGmVF
One CT scan, one patient graph. Our lung nodule study connects slice features and learns attention weights for their relationships, evaluating malignancy classification on LIDC-IDRI.
#MedicalImaging#GraphNeuralNetworks
https://t.co/LMfoqAvo5d
Does every layer need the same precision? Our FPGA ECG classifier uses configurable quantization-aware training to assign precision layer by layer in a streaming Inception-ResNeXt accelerator.
#ECG#FPGA
https://t.co/XFw1s2uymy
Matching EEG feature distributions can blur class boundaries. Our preprint adds class-specific prototype alignment to global domain adaptation, evaluated across subjects in MODMA.
#EEG#DomainAdaptation
https://t.co/NSg6eDWeJH
What happens to image enhancement when the hardware varies? Our handheld photoacoustic study models FeRAM crossbar variation alongside MultiResU-Net, using simulated and in-vivo image data.
#Photoacoustics#FeRAM
https://t.co/HCUyTU6asJ
A shared radio has to make time for synchronization. In our wireless timer study, lighter BLE traffic and longer connection intervals improve timing accuracy by reducing contention.
#IoT#WirelessCommunication
https://t.co/Ra3H82FMPb