Get Started
Topic · 4 recaps
Small / On-device Models
Compact language and vision models designed to run on phones, laptops, and edge hardware — usually under a few billion parameters, often with aggressive distillation or quantization.
Play all
...
Posts
Questions
Home
Topics
Search
Library
Questions researchers are working on
Follow a question through Rcap’s explanations and the latest papers addressing it.
Search
Can post-training ternarization make language models smaller without unacceptable capability loss or slower inference?
Ultra-low-bit weights can shrink model storage, but nominal bit counts may not reflect the stored representation, uneven task degradation, or actual inference speed. Compression may therefore improve footprint without improving end-to-end deployment performance.
Can preprocessing defenses detect adversarial attacks in depthwise-separable edge vision CNNs when they cannot restore predictions?
Preprocessing defenses are often assumed to transfer across model architectures, but depthwise-separable CNNs may respond differently to adversarial perturbations than residual or Inception-style networks. Their failure to recover predictions may still produce measurable differences between clean and adversarial inputs, while image-quality scores may not reflect defensive value.
How can an accessible humanoid robot integrate multimodal AI for real-world human interaction?
Real-world human-robot interaction requires coordinating visual, gestural, and spoken inputs with physical manipulation. Integrating these capabilities on an accessible humanoid platform also requires maintaining accurate, timely control across the system.
How can audio deepfake detectors identify and localize manipulation when genuine and fake content coexist?
A whole-clip label can conceal which time intervals or overlapping sources provide evidence of manipulation. This makes mixed-authenticity audio decisions difficult to interpret and verify.
How can auscultation waveforms detect arteriovenous fistula dysfunction robustly across patients on resource-constrained devices?
Arteriovenous fistula dysfunction must be detected from sound recordings despite patient-specific variation and limited device compute. Conventional feature extraction may not transfer reliably across patients or after dimensionality reduction.
How can automated quality control reliably grade artifact severity in ultra-low-field neonatal brain MRI with practical inference costs?
Low signal-to-noise ratio, absent shielding, and long scans make ultra-low-field neonatal brain MRI vulnerable to acquisition artifacts. Quality control must distinguish severity across several artifact types while remaining practical to deploy.
How can CNN inference fit on bare-metal microcontrollers with severe memory, binary-size, and runtime limits?
Bare-metal microcontrollers may provide only a few kilobytes of SRAM and limited firmware space, while CNN inference also imposes execution-time demands. These constraints can make standard CNN deployment impractical even when the model's accuracy is acceptable.
How can compact multimodal Earth-observation models handle missing sensors and changing spatial resolutions?
Earth-observation systems often receive incomplete combinations of heterogeneous sensor data, while downstream imagery may use resolutions different from pretraining. The challenge is to preserve useful cross-sensor representations without relying on a large parameter budget.
How can compact Persian medical QA models reason reliably and estimate answer confidence on consumer hardware?
Persian medical QA remains underserved, and compact models must handle clinical reasoning despite limited language-specific resources. Reliable confidence estimates are also needed to distinguish answers that may be unsafe to use.
How can compact text embedding models improve retrieval and generalization through better training and data quality?
Compact embedding models can underperform on retrieval and generalization when development emphasizes data scaling or synthesis without sufficiently addressing training methods and data quality. The problem is to improve their representations without relying on substantially larger models.
How can complex neural-network graphs be mapped across heterogeneous SoCs to balance inference latency and throughput?
Complex neural-network graphs expose parallel operators, but executing them across unlike on-chip processors introduces dependencies and communication costs. Choices that improve pipeline throughput can worsen single-inference latency or energy efficiency.
How can depth compression merge padded convolutions without enlarging kernels for faster embedded inference?
Padding complicates analytical layer merging, while enlarged kernels can offset the speed benefits of reducing network depth. The challenge is to merge layers efficiently without increasing kernel size.
How can diffusion language models support reliable mobile-edge agents under tight latency and resource constraints?
Mobile-edge agents must generate and act despite limited compute, memory, energy, and network capacity while meeting privacy and reliability requirements. Diffusion language models refine multiple tokens with bidirectional context, creating latency and quality trade-offs that differ from sequential decoding.
How can document parsing remain accurate on difficult pages while running efficiently on limited GPUs?
Document parsing must recover text, formulas, tables, and structure from varied pages, but complex layouts can increase errors and decoding cost. These costs make accurate OCR harder to deploy on modest hardware.
How can edge analog in-memory inference reduce energy while certifying errors from heterogeneous, imperfect accelerators?
Analog in-memory accelerators save energy by computing within memory, but device faults, programming errors, noise, and converter limits distort predictions differently across chips. Activating multiple accelerators can improve reliability at substantial energy cost, while relying on one may leave its error rate uncertain.
How can environmental-audio generators provide semantic control under limited compute and training data?
Generating everyday sounds with meaningful semantic control can require substantial computation and large, carefully curated datasets. These requirements make controllable environmental-audio synthesis difficult when training resources and data are limited.
How can Gaussian splatting reduce per-primitive appearance memory without limiting angular detail?
In 3D Gaussian Splatting, spherical-harmonic coefficients consume substantial per-primitive storage and memory traffic, while their band-limited basis can miss fine view-dependent effects. More expressive representations can also change optimization behavior and absorb non-static scene content, complicating the quality–efficiency trade-off.
How can image restoration remove varied weather degradations from high-resolution images under tight compute budgets?
Rain, haze, and snow can obscure image content, while high-resolution inputs make restoration computationally expensive. A single restoration system must address these different degradations without exceeding available computing resources.
How can image-based deepfake detectors stay accurate on low-resolution inputs under tight compute budgets?
Low-resolution inputs can reduce the visual evidence available for detecting manipulation, while expensive inference can prevent deployment in resource-constrained environments.
How can image-to-shape diffusion transformers be compressed for resource-constrained 3D generation without losing geometric fidelity?
Large diffusion transformers make 3D shape generation difficult in resource-constrained environments. Compression strategies from other domains may fail to account for the unequal geometric importance of different transformer components.
Previous
1 / 4
Next