Direct Inference on Compressed Measurements

32 concepts28 papers2 proposals

In traditional signal processing and computer vision pipelines, data compression and semantic inference live on opposite sides of a massive computational wall. Sensors capture high-dimensional signals, compress them for storage or transmission, and then—before any analysis can happen—a computer performs an expensive inverse operation to reconstruct the original pixels or waveforms.

This "reconstruct-then-analyze" pipeline introduces a severe bottleneck. Reconstructing a compressively sensed image or signal requires solving an underdetermined system of equations, typically relying on slow, iterative optimization algorithms. As Kuldeep Kulkarni et al. note in ReconNet: Non-Iterative Reconstruction of Images from Compressively Sensed Measurements, this intermediate reconstruction phase introduces significant latency.

But if the ultimate goal of your system is to make a decision—such as detecting an anomaly, classifying an object, or diagnosing a disease—is full signal reconstruction actually necessary? A growing body of research shows that it is not. By performing inference directly within the compressed domain, we can bypass the reconstruction step entirely, saving massive amounts of compute, power, and time.

The Intuition: Silhouette vs. Portrait

To understand why direct compressed-domain inference works, consider a simple analogy: recognizing a silhouette versus painting a high-fidelity portrait.

[High-Dimensional Signal] ---> [Linear Projection / Sketching] ---> [Compressed Domain]
                                                                            |
       +--------------------------------------------------------------------+
       | (Traditional Path)                                                 | (Direct Path)
       v                                                                    v
[Iterative Reconstruction] (Heavy Compute)                          [Direct Inference] (Light Compute)
       |                                                                    |
       v                                                                    v
[High-Fidelity Pixels] ---> [Downstream Classifier]                 [Semantic Label / Decision]

Reconstructing a signal means finding its exact coordinates in a high-dimensional space. This requires a high volume of measurements to resolve ambiguities. Classification, however, only requires determining which side of a decision boundary a signal falls on. The amount of information needed to find a coarse semantic boundary is fundamentally lower than the information needed to rebuild every pixel.

Hugo Reboredo et al. formalized this mathematical intuition in Bounds on the Number of Measurements for Reliable Compressive Classification. By modeling signal classes as low-dimensional Gaussian subspaces, the authors proved that reliable classification requires significantly fewer measurements than full signal recovery. The geometry that separates different classes remains intact even after severe dimensionality reduction.

Because this geometry is preserved, standard decision rules can be mapped directly into the compressed space. For example, Hugo Reboredo et al. demonstrated in Compressive classification that Maximum A Posteriori (MAP) classification can operate directly on low-dimensional projections. Similarly, Junni Zou et al. showed in Compressive detection with sparse random projections that by projecting a known signal template into the measurement space, one can perform hypothesis testing via a sufficient statistic derived entirely from compressed samples. This completely eliminates the need for complex $\ell_1$-minimization algorithms. Kalyani Krishnamurthy et al. confirmed in Target detection performance bounds in compressive imaging that such direct decision boundaries yield reliable target and anomaly detection while bypassing computational latency.

Preserving Geometry Without Decompression

How do machine learning models compute similarities or find boundaries when the underlying data is compressed? The secret lies in the fact that linear compressions—like random projections or sketching—act as geometric preserves.

Kernel and Distance Approximations

Many classification algorithms rely on computing inner products or distances between data points. In Random Features for Sparse Signal Classification, Aswin C. Sankaranarayanan et al. introduced Compressive Random Features (CRF), proving that kernel similarities of high-dimensional signals can be accurately approximated directly from their compressed measurements. This allows kernel-based classifiers to operate efficiently in the compressed domain, scaling with signal sparsity rather than raw pixel count.

Similarly, for massive multi-label problems, Xiuwen Gong et al. utilized stochastic sketching in Fast Multi-label Learning to compress data into compact "sketches," performing k-nearest neighbor searches directly within the compressed embedding space to preserve local data structures.

Structural and Quantized Learning

This preservation holds true even when dealing with structured matrix data or extreme quantization:

End-to-End Co-Design of Hardware and Networks

While random mathematical projections provide strong theoretical guarantees, modern deep learning allows us to learn optimal, task-specific compression matrices. By treating the physical sensing mechanism as the first layer of a neural network, hardware and software can be co-designed to capture only the features necessary for inference.

[Physical Scene] ---> [Optimized Optical Mask / Sensor] ---> [Compressed Latent Representation] ---> [Deep Inference Network] ---> [Label]
                             ^                                              |
                             +------------------ Jointly Optimized ---------+

In Deep Coding Patterns Design for Compressive Near-Infrared Spectral Classification, Alejandra Hernandez-Rojas et al. implemented an end-to-end framework where a camera's physical optical coding patterns are optimized jointly with a classification network. The compressed measurements act as an optimized latent representation, proving that compressed signatures retain sufficient embedded information for deep learning models to execute accurate inference without data cube restoration.

Similarly, Henry Arguello et al. applied this concept to 3D vision in Fast Disparity Estimation from a Single Compressed Light Field Measurement. By mapping a single coded 2D snapshot directly to 3D scene geometry, the authors bypassed light field reconstruction entirely, drastically reducing inference latency and error propagation.

Real-World Impact: Edge Devices and Diagnostics

Bypassing reconstruction provides massive practical advantages in environments constrained by power, bandwidth, or time.

Wearables and Intelligent Vision

High-fidelity signals create severe data bottlenecks on tiny devices. To solve this, Di Wu et al. developed C2SP-Net: Joint Compression and Classification Network for Epilepsy Seizure Prediction, an end-to-end framework that compresses EEG signals and performs seizure prediction directly on the compressed data. This eliminates the need for power-hungry signal reconstruction at the edge, reducing computational overhead by roughly 14 times.

Taking this a step further, Haoxin Huang et al. moved compression directly into hardware in In-sensor compressing via programmable optoelectronic sensors based on van der Waals heterostructures for intelligent machine vision. Their programmable optoelectronic sensors retain critical semantic information directly at the sensing site, cutting downstream model parameters by nearly 70% and bypassing the memory wall associated with raw data transmission.

Video Streams and Medical Diagnostics

Direct inference also accelerates software-level pipelines where data is already compressed:

  • Video Bitstreams: In Can you tell a face from a HEVC bitstream?, Saeed Ranjbar Alvar et al. demonstrated that face detection can be performed by training shallow CNNs directly on raw syntax features from an entropy decoder, skipping over 60% of the HEVC decoding process and running 2–3 times faster than pixel-domain methods.
  • Image-less MRI Diagnostics: In medical imaging, generating human-interpretable scans is often a slow, artifact-prone detour if the end goal is simply a medical diagnosis. Addressing this, Yuning Du et al. introduced an image-less active sampling framework in The MRI Scanner as a Diagnostic: Image-less Active Sampling. Using reinforcement learning to sample raw, undersampled k-space data dynamically, their system achieves high diagnostic accuracy without ever reconstructing an image. When reconstruction is eventually desired, compressed-domain categorization can still assist; Yinghua Li et al. showed in Super-Resolution of Brain MRI Images Using Overcomplete Dictionaries and Nonlocal Similarity that classifying image patches directly within the compressed frequency domain improves both the efficiency and accuracy of subsequent super-resolution pipelines.

By recognizing that semantic information can be extracted directly from low-dimensional projections, reconstruction-free inference fundamentally redefines the relationship between sensing and computing, proving that we can understand the world through data we never fully rebuild.

Key Foundational Work

The smashed filter for compressive classification and target recognition (2007) — contributed 2 concept elements to this direction.
Target detection performance bounds in compressive imaging (2012) — contributed 2 concept elements to this direction.
Compressive Sampling for Signal Detection (2007) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_reconstruction_free_classification_e148c288,
  title = {reconstruction-free classification},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/reconstruction-free-classification/rdr_e148c288aed842f4bdf5122c98f7e375},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_reconstruction_free_classification_e148c288_v1,
  title = {reconstruction-free classification},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/reconstruction-free-classification/rdr_e148c288aed842f4bdf5122c98f7e375/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake