Pith. sign in

REVIEW 4 major objections 6 minor 127 references

Semantic Communication based on Generative AI: A New Approach to Image Compression and Edge Optimization

T0 review · 4 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This thesis argues that transmitting a semantic segmentation map and letting a generative model reconstruct the image preserves meaning at roughly a tenth of the bitrate that classical codecs need.

desk verdict Plausible semantic compression results, but the headline bitrate gains rest on a self-referential metric; worth serious refereeing if the evaluation is made independent. read the letter →

arxiv 2502.01675 v1 pith:ROMGCZ4N submitted 2025-02-01 cs.CV cs.AIcs.LG

Francesco Pezone This is my paper · ORCID
classification cs.CVcs.AIcs.LG
keywords SemanticcommunicationGenerativeimagecompressionDiffusionmodelsVectorquantizationsegmentationmapsGoal-orientedEdgenetworkresourceallocationInformationBottleneck
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This thesis makes the case that image communication should transmit meaning rather than pixels, and that generative models can turn that principle into working codecs. It proposes three semantic-preserving coding schemes—SPIC, C-SPIC, and SQ-GAN—that compress images by sending a semantic segmentation map plus a low-resolution or masked vector-quantized side signal, then letting a generative model at the receiver reconstruct the image. The central claim, supported on the Cityscapes benchmark, is that these schemes preserve semantic content—measured by mean intersection-over-union, FID, and traffic-sign classification accuracy—at bitrates roughly an order of magnitude lower than classical codecs such as BPG and JPEG2000. If true, the practical payoff is that autonomous driving and other real-time visual tasks could run over far leaner links with the same semantic reliability, and edge networks could dynamically trade bandwidth against reconstruction quality and task accuracy. The thesis also embeds the codecs in a goal-oriented resource-allocation framework that minimizes power consumption subject to delay and task-performance constraints.

What carries the argument

The central object is the semantic segmentation map (SSM), a per-pixel class label tensor that both transmitter and receiver treat as the shared knowledge base. The mechanisms that carry the argument are: SemCoRe, a U-Net diffusion model doubly conditioned on the lossless SSM through SPADE (spatially-adaptive normalization) blocks and on the upscaled coarse image through channel concatenation; the Semantic-Conditioned Adaptive Mask Module (SAMM), which gives every latent vector a relevance score conditioned on the SSM and quantizes only the top-scoring vectors; the class-masked residual pipeline of C-SPIC, which compacts and rescales the difference between the original and the intermediate reconstruction over selected classes; and a goal-oriented edge optimizer built on the Information Bottleneck principle that adjusts compression and transmission resources. Masking fractions and coarse/residual bitrates are the explicit knobs that let the same codec operate across a rate-quality continuum.

What would settle it

Reconstruct a batch of images with SPIC, C-SPIC, or SQ-GAN and compare the overlap score computed with the same segmentation model used at the transmitter against the score computed with an independently trained segmenter, while also auditing the reconstructions for small or rare objects that the transmitted map may have missed. If the second score drops while the first stays high, the semantic-preservation claim is an artifact of sharing the labeler.

Watch

Extended reading notes

Core claim

The result the author is trying to establish is that a semantic segmentation map can act as a near-sufficient carrier of image meaning: send the map, plus a small side signal, and a generative model can restore an image whose semantic content—object classes, layout, small detailed objects—survives at bitrates where classical codecs destroy it. Concretely, SPIC transmits the segmentation map losslessly and a downscaled coarse image lossily; SemCoRe then runs a diffusion reverse process conditioned on both. In the reported Cityscapes experiments, SPIC at 0.17 bits per pixel beats BPG at 1 bit per pixel on mIoU, and C-SPIC adds a compacted, class-masked residual so that a traffic sign misrendered at 0.166 BPP becomes readable at 0.171 BPP. SQ-GAN replaces the coarse image with masked vector-quantized codes whose masking fractions are set by semantic relevance, and the thesis reports it outperforming BPG and JPEG2000 on both classical and semantic metrics at extremely low BPP. The broader thesis-level claim is that these per-scene bitrate savings can be converted into network-level savings through a goal-oriented resource-allocation layer that minimizes power while respecting delay and task-accuracy constraints.

Load-bearing premise

The load-bearing premise is that the semantic segmentation map produced by the transmitter's off-the-shelf model contains all the meaning the receiver needs; if the map misses or mislabels an object, the reconstruction permanently loses it, and an overlap-based segmentation evaluation cannot detect the loss when the same kind of model labels the output.

Editorial extensions

If this is right

  • Semantic image compression can preserve object layout and class content at bitrates where classical codecs destroy them, so machine-vision links can run on a fraction of the bandwidth.
  • Small, safety-critical objects that a coarse side signal loses can be recovered by sending a compacted residual for selected classes at a small bitrate overhead, around 0.004–0.005 BPP in the reported traffic-sign examples.
  • Because masking fractions and coarse/residual rates are adjustable in SQ-GAN and C-SPIC, the same codec can be throttled in real time to meet network constraints, enabling goal-oriented edge resource allocation.
  • The modular structure lets the segmentation model, lossless codec, or lossy codec be swapped without retraining the generative receiver, so improvements in off-the-shelf components transfer directly.
  • Combined with the Information Bottleneck-based allocation, the schemes give a way to minimize average power consumption while keeping average delay and task metrics within constraints.
  • The same architecture could be pointed at other compact structured side information—depth maps, object bounding boxes, or pose graphs—wherever a small signal captures what a downstream task needs.
  • An implicit scheduling rule suggested by the design is to allocate residual bits to classes by their contribution to the task rather than by pixel area; the C-SPIC results show this can be nearly free in total bitrate.
  • A natural stress test, not run in the thesis, is to re-score reconstructed images with a segmentation model different from the one that produced the transmitted map; that would separate true semantic preservation from consistency with the transmitter's own labeler.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The thesis proposes a semantic-communication framework for image compression and edge-network optimization. Chapter 4 introduces SPIC, a modular scheme in which the transmitter sends a losslessly compressed semantic segmentation map (SSM) plus a low-resolution coarse image, and a receiver-side diffusion model (SemCoRe) reconstructs a high-resolution image conditioned on both. C-SPIC variants add a class-specific residual for semantically relevant objects. Chapter 5 introduces SQ-GAN, a VQ-GAN variant with SSM-conditioned masking (SAMM), a semantic positional embedding, and a semantic-aware discriminator. Chapter 6 formulates goal-oriented resource allocation in an edge network using the Gaussian Information Bottleneck and stochastic optimization, and integrates SQ-GAN into the framework. The central empirical claim is that SPIC, C-SPIC, and SQ-GAN preserve semantic metrics (mIoU, FID, traffic-sign classification accuracy) at bitrates roughly an order of magnitude lower than BPG and JPEG2000 on the Cityscapes dataset.

Significance. If the empirical claims hold, the thesis is a useful demonstration that generative priors combined with task-oriented side information can move the rate-semantic-distortion frontier for machine-consumed imagery. The writing is clear, the architectural descriptions are detailed, and the comparison against BPG and JPEG2000 is a sensible baseline choice. The work also makes the modularity argument concretely: SPIC can be extended to C-SPIC without retraining the diffusion model, and SQ-GAN exposes explicit masking fractions for rate control. However, the manuscript does not ship code or trained models, and the quantitative evidence is limited to point estimates without error bars. More importantly, the main evaluation metrics are closely aligned with the very side information that the transmitter sends, so the magnitude of the claimed gains needs independent verification before the order-of-magnitude bitrate advantage can be accepted.

major comments (4)
  1. [§4.2.1, §4.4.1, Fig. 4.7] The mIoU metric in Fig. 4.7 is evaluated between the SSM produced by INTERN-2.5 on the original image and the SSM produced by the same INTERN-2.5 on the reconstructed image. Because the transmitted SSM is exactly the INTERN-2.5 output on the original image and SemCoRe is conditioned on that SSM, this mIoU largely measures how well the decoder re-expresses its conditioning input, not independent semantic preservation. The red upper bound at mIoU 0.84, being INTERN-2.5's self-consistency on original images, confirms that the metric is internal to the chosen segmenter. The order-of-magnitude bitrate advantage claimed from this figure should be supported by an evaluation using a different segmentation model than the one providing the side information, and by task metrics that do not derive from the transmitted SSM.
  2. [§3.6, §4.4.2, Fig. 4.9] The traffic-sign classification accuracy metric localizes signs using the original SSM and then classifies the corresponding crops of the original and reconstructed images. C-SPIC w transmits a class-specific residual for exactly the traffic-sign class, so a high ACC at low BPP is partly a direct consequence of the transmitted side information rather than evidence of general semantic preservation. The metric also cannot detect semantic loss for traffic signs that the SSM failed to localize in the first place. A more stringent test would be to localize signs from the reconstruction itself or to use a detector independent of the transmitted SSM.
  3. [§4.4, Figs. 4.7 and 4.9] All quantitative comparisons are point estimates averaged over 500 images without error bars, confidence intervals, or statistical tests. The reported advantages are sometimes modest in absolute terms (e.g., SPIC at 0.17 BPP versus BPG at roughly 1 BPP on mIoU), and without variance information the central order-of-magnitude claim is not fully established. Please report per-seed variability or bootstrap confidence intervals, and state the number of independent runs used for each operating point.
  4. [§5.4, Fig. 5.11] For the SQ-GAN results it is unclear whether the SSM used at test time is the ground-truth Cityscapes annotation or an SSM extracted by a segmentation model. Since the architecture compresses the SSM, using ground-truth annotations at test time would omit the cost of producing the SSM at the transmitter and would make the comparison with BPG not deployment-equivalent. Please specify the exact test-time SSM source, the procedure for matching BPP between SQ-GAN and the classical codecs, and how the masking fractions mx and ms were selected for each rate point.
minor comments (6)
  1. [§4.4.2, last paragraph] The phrase 'glssc-based image compression algorithms' appears to be a typo; it should likely read 'semantic-based' or 'generative-based'.
  2. [§1.3, throughout] The term 'out-of-the-shelf' is used repeatedly; the standard phrasing is 'off-the-shelf'.
  3. [§4.4.2, Fig. 4.9] The three shades of blue used for the C-SPIC w curves are difficult to distinguish in print; adding distinct markers or line styles would improve readability.
  4. [§5.2.1] The sentence 'The presence of the apex/subscripts x and s will be used to refer to the specific pipeline' is awkward; consider rephrasing to 'Subscripts x and s denote the image and SSM pipelines, respectively.'
  5. [§3.6, ACC definition] The notation 'x ← → ˆx' used in the ACC definition is introduced only in Chapter 2; a brief cross-reference would help the reader.
  6. [§4.2.2] The inference procedure mentions T = 20 diffusion steps but does not specify whether DDIM or the standard DDPM sampling schedule is used; this detail is relevant for reproducibility.

Circularity Check

1 steps flagged · score 4.0 of 10

Semantic metric is self-referential: the same INTERN-2.5 model that produces the transmitted SSM also scores the reconstructed image, so part of the headline mIoU bitrate gain is built into the evaluation setup; the rest of the thesis is an independent empirical study.

  1. self definitional [Sections 4.2.1 and 4.4.1, Fig. 4.7; SSM definition in Section 1.3]
    "In this work, the SSM s is extracted directly from the original image x using an out-of-the-shelf pre-trained SOTA SS-Model. The INTERN-2.5 model [116] is chosen for this task due to its high performance in semantic segmentation. ... This quantity is evaluated by the mIoU between the original SSM and the one generated from x via the INTERN-2.5 SS-Model."

    The transmitter defines the semantic content as s = INTERN-2.5(x) and sends it losslessly; the receiver generates x_hat = SemCoRe(s, c), a diffusion model conditioned and trained to preserve s. The headline semantic metric is mIoU(s, INTERN-2.5(x_hat)), i.e., the same extractor is applied to a reconstruction that was conditioned on that extractor's own output. High mIoU therefore measures self-consistency with the conditioning input rather than independent semantic preservation, while BPG is scored without receiving any such semantic side information. The claimed order-of-magnitude bitrate advantage on mIoU is partly enforced by the evaluation setup rather than discovered.

full rationale

The central SPIC/C-SPIC contribution is an empirical compression study rather than a first-principles derivation, so most of the paper does not reduce to its inputs. The mIoU comparison in Fig. 4.7 is the one load-bearing step that is circular by construction: the same INTERN-2.5 segmenter supplies the transmitted SSM (Sec. 4.2.1) and grades the reconstruction (Sec. 4.4.1), with the red 0.84 mIoU line explicitly acknowledging the model's self-bound. However, this is only a partial circularity. The thesis also offers independent evidence (FID using Inception-v3 features, masked PSNR, and qualitative visual comparisons) that does not rely on the same segmentation model, and the edge-optimization chapter uses the externally derived Gaussian Information Bottleneck solution of Chechik et al. rather than a self-citation chain. The self-citations to the author's ICASSP papers are real published studies and are not used as unverified uniqueness arguments. Overall, one self-referential metric inflates the headline semantic bitrate gain, but the central claims retain independent content, so the circularity score is moderate.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical entities (particles, forces, dimensions) are postulated. The proposed modules (SemCoRe, SAMM, SemPE) are architectural components within known machine learning frameworks, not entities requiring independent evidence outside the paper.

free parameters (3)
  • SQ-GAN masking fraction mx (and ms) = adjustable between 0 and 1; used in results curves
    This ratio determines how many latent vectors are kept after semantic masking; the paper reports results across several masking fractions, so it is an operating point rather than a fitted constant.
  • BPG quality level for SPIC coarse image = e.g., 0.007, 0.054, 0.183 BPP in C-SPIC experiments
    Chosen to explore rate-distortion trade-offs; the central claims depend on the coarse channel bitrate.
  • Number of diffusion steps at inference (T) = 20 (vs 1000 training steps)
    Selected to balance compute and reconstruction fidelity; not derived from first principles.
assumptions (4)
  • standard math Gaussian Information Bottleneck provides the optimal linear encoder for jointly Gaussian x and y (Chechik et al. 2005)
    Used in Section 6.3 for edge resource allocation.
  • domain assumption A semantic segmentation map (SSM) captures the semantic information of an image
    This is the foundational premise of SPIC and SQ-GAN; if false, semantic preservation fails.
  • domain assumption A pre-trained segmentation model (INTERN-2.5) generates accurate SSMs on Cityscapes-like data
    Used at inference in Sections 4.2.1 and 5.1; errors in the SSM propagate to reconstruction.
  • ad hoc to paper Lossless compression of the SSM with FLIF achieves about 0.112 BPP on Cityscapes
    This empirical value is measured on Cityscapes and is used to compute total BPP; it may not generalize.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Semantic Communication based on Generative AI: A New Approach to Image Compression and Edge Optimization." pith.science (2026). https://pith.science/paper/ROMGCZ4N

@misc{pith2026250201675,
  author       = {Pith},
  title        = {Pith review of: Semantic Communication based on Generative AI: A New Approach to Image Compression and Edge Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ROMGCZ4N}},
  note         = {Machine review of arXiv:2502.01675}
}
read the original abstract

As digital technologies advance, communication networks face challenges in handling the vast data generated by intelligent devices. Autonomous vehicles, smart sensors, and IoT systems necessitate new paradigms. This thesis addresses these challenges by integrating semantic communication and generative models for optimized image compression and edge network resource allocation. Unlike bit-centric systems, semantic communication prioritizes transmitting meaningful data specifically selected to convey the meaning rather than obtain a faithful representation of the original data. The communication infrastructure can benefit to significant improvements in bandwidth efficiency and latency reduction. Central to this work is the design of semantic-preserving image compression using Generative Adversarial Networks and Denoising Diffusion Probabilistic Models. These models compress images by encoding only semantically relevant features, allowing for high-quality reconstruction with minimal transmission. Additionally, a Goal-Oriented edge network optimization framework is introduced, leveraging the Information Bottleneck principle and stochastic optimization to dynamically allocate resources and enhance efficiency. By integrating semantic communication into edge networks, this approach balances computational efficiency and communication effectiveness, making it suitable for real-time applications. The thesis compares semantic-aware models with conventional image compression techniques using classical and semantic evaluation metrics. Results demonstrate the potential of combining generative AI and semantic communication to create more efficient semantic-goal-oriented communication networks that meet the demands of modern data-driven applications.

Figures

Figures reproduced from arXiv: 2502.01675 by the authors.

Figure 1.1
Figure 1.1. Diagram illustrating the three lev￾els of communication and their interconnection as proposed in [117] This idea of going beyond the classical bit-by-bit reconstruction was proposed by Weaver in 1953 when he introduced the three different levels of communi￾cation [117]: (i) the syntactic level, focusing on the accurate transmission of symbols; (ii) the semantic level, focusing on the conveyance of meaning; and (iii)… view at source ↗
Figure 2.1
Figure 2.1. Semantic Communication scheme. 2.3 Semantic Communication (Semantic Level) On the semantic level, the way the message is reconstructed is not relevant as long as the semantic information is preserved. The term semantic information refers to the information that is conveyed by the data and is relevant to the receiver, allowing the receiver to understand the message without reconstructing it symbol by symbol. Multiple… view at source ↗
Figure 2.2
Figure 2.2. Impact of varying channel conditions on semantic reconstruction and how the generative model adapt to these changes. As the channel quality improves, the generative model is able to receive more semantic symbols zi and eventually reach semantically equivalence between x and xˆ. As soon as the channel conditions improve, the generative model will be able to receive more semantic symbols zi and generate images that ar… view at source ↗
Figures from the paper (37 more)
Figure 2.3
Figure 2.3. Figure 2.3: Goal-Oriented Communication Scheme. limitations. In this context consider y as the data that contains the necessary information to perform a certain task at the receiver. The idea of GOC is to compress and transform x to obtain a representation z = Φ(x) that is able …
Figure 2.4
Figure 2.4. Figure 2.4: Trade-off representation of the IB problem in the Relevance-Complexity plane [122]. The blue line represents the optimal solutions for different values of β while the red line represents the limit of relevance given by all the information that x has on y. The optimal…
Figure 2.5
Figure 2.5. Figure 2.5: Semantic-Goal-Oriented Communication Scheme. and force it to adjust the value of β. By modifying the complexity of the transformation and its relevance, this approach becomes a powerful tool for facilitating the optimization of communication. 2.5 Semantic-Goal-Orient…
Figure 3.1
Figure 3.1. Figure 3.1: Architectural diagram of the ResBlock. The residual connection is rep￾resented by the line that goes from hl−1 to the sum symbol. • Encoder: The encoder network is the interface between the input data x and the latent representation z. Its goal is to reduce the compl…
Figure 3.2
Figure 3.2. Figure 3.2: Architectural diagram of the U-Net as proposed in [82]. The encoder and decoder are linked via the skip connections represented in gray. between the encoder and decoder. These additional layers are called bottleneck layers and are used to further process the latent r…
Figure 3.3
Figure 3.3. Figure 3.3: Architectural diagram of the ResBlock conditioned on the time t. The value t is mapped to an higher dimensional space via the Positional Embedding (PE) before being inserted in the ResBlock. advantage: (i) during backpropagation, they allow the gradient to flow from …
Figure 3.4
Figure 3.4. Figure 3.4: On the left, the architectural diagram of the ResBlock z proposed in [73] that introduces SPADE, a normalization technique able to condition on the SSM . On the right, the SPADE architecture is compared with batch normalization. The parameters γ and β are now learned…
Figure 3.5
Figure 3.5. Figure 3.5: Overview of the VAE architecture. The latent tensor is obtained as z = µ + σ ⊙ ϵ via the reparameterization trick by selecting a random value ϵ ∼ N (0, I). q(z|x). They represent the core processes of encoding the data x to a latent representation z and then retrievi…
Figure 3.6
Figure 3.6. Figure 3.6: Overview of the GAN architecture. The example reported involves a SR task where the Gener￾ator tries to up-scale the xLR such that the reconstructed xˆ can has enough details to fool the Discriminator. In 2014, Goodfellow et al. introduced the GAN [25]. Unlike previo…
Figure 3.7
Figure 3.7. Figure 3.7: Overview of the VQ-VAE architecture. After the encoding the latent vectors zk are vector￾quantized via the learnable codebook C. The decoder then uses the quantized latent tensor zq to generate the output instead of the original z. An interesting variant of the class…
Figure 3.8
Figure 3.8. Figure 3.8: Overview of the VQ-GAN architecture as proposed in [22]. Similar to the VQ-VAE the latent vectors are vector-quantized and used in the decoding process. However, only the first N codes from the top-left corner to the bottom-right are used while the other K − N are pr…
Figure 3.9
Figure 3.9. Figure 3.9: Overview of the MQ-VAE architecture as proposed in [38]. Differently form the VQ-GAN now the introduction of the Adaptive Mask Module select the relevant latent vectors zk in order of importance instead that in order of appearance. The Adaptive De-Mask Module is used…
Figure 2
Figure 2. Figure 2: The Illustration of our proposed two-stage generation framework. (a) In the first stage, MQ-VAE adaptively masks the redundant quantizes only those regions that are most critical for reconstruction. This selection is performed [PITH_FULL_IMAGE:figures/full_fig_p044_2.png]
Figure 3.10
Figure 3.10. Figure 3.10: Visual representation of the DDPM forward and reverse process [34]. The forward process (from right to left) identified by q(xt|xt−1) is implemented by gradually adding Gaussian white noise to the original image. Starting from the x0 at the end of T iterations the r…
Figure 3.11
Figure 3.11. Figure 3.11: The control-flow graphical scheme used for the CFG mechanism as proposed in [35]. The same U-Net model is trained with and without conditioning to allow control over the output. This method allows for flexible and controlled generation. If the dataset contains multi…
Figure 4.1
Figure 4.1. Figure 4.1: Overview of the SPIC architecture. At the transmitter side the SSM s and the coarse image c are extracted from the original image and compressed with classical off-the-shelf compression algorithms. At the receiver it is employed the SemCoRe that leverages both s and …
Figure 4.2
Figure 4.2. Figure 4.2: Overview of the C-SPIC w architecture. At the transmitter after the generation and compression of s and c the SemCoRe is used to evaluate the intermediate reconstructed image x˜ and the relevant residual r. This residual is processed and compacted to preserve only in…
Figure 4.3
Figure 4.3. Figure 4.3: Processing pipeline of the residual r in the C-SPIC : (a) the residual isis initially masked using the binary mask m, (b) the function g(·) is applied to compact the relevant regions and reduce empty spaces, and (c) the pixel values are re-scaled to an 8-bit represen…
Figure 4.4
Figure 4.4. Figure 4.4: Overview of the C-SPIC w/o architecture. At the transmitter after the generation and com￾pression of s and c the original image is directly used to extract the information about relevant classes. After masking and compacting, the transformation h ′ is send to the rec…
Figure 4.5
Figure 4.5. Figure 4.5: Visual comparison between the original image and SSM (CENTER), the image compressed with BPG at 0.176 BPP with the associated generated SSM (LEFT), and the image obtained with the proposed SPIC framework at 0.166 BPP and the associated generated SSM (RIGHT). At simil…
Figure 4.6
Figure 4.6. Figure 4.6: Visual comparison between: the SOTA SR model [81] (LEFT), the proposed SPIC framework (CENTER), and the SSM conditioned model [115] (RIGHT). larger foreground objects the level of detail and SSM retention is good enough. However, as the object size shrinks, conventio…
Figure 4.7
Figure 4.7. Figure 4.7: Performance comparison between BPG, JPEG2000 and SPIC evaluated in terms of mIoU and FID. only on the SSM using the model proposed in [115]. In this case, the pedestrians are placed in the correct positions, as are all the other elements in the scene. Still, the draw…
Figure 4.8
Figure 4.8. Figure 4.8: Visual comparison of the traffic sign "P" compressed with the BPG algorithm at 0.176 BPP (LEFT), the proposed SPIC framework at 0.166 BPP(CENTER), and the C-SPIC w at 0.171 BPP (RIGHT). BPP PSNR (a) BPP ACC (b) [PITH_FULL_IMAGE:figures/full_fig_p070_4_8.png]
Figure 4.9
Figure 4.9. Figure 4.9: Performance comparison between BPG, JPEG2000 and C-SPIC w on the "traffic sign" class evaluated in term of masked PSNR and traffic signs classification accuracy. In [PITH_FULL_IMAGE:figures/full_fig_p070_4_9.png]
Figure 5.1
Figure 5.1. Figure 5.1: Overview of the generator of the SQ-GAN architecture. The generator is composed of two sub-networks: Gs in beige and Gx in green. They are responsible for the compression and reconstruction of the SSM and the image, respectively. number of channels and H16 and W16 ar…
Figure 5.2
Figure 5.2. Figure 5.2: Architectural diagram of the AMM as in [38] (LEFT), and the proposed SAMM employing the SPADE layer to introduce the SSM conditioning (RIGHT). 5.2.2 Image Encoder At the same time, the image x is processed by the image encoder Ex. This block is responsible for mappin…
Figure 5.3
Figure 5.3. Figure 5.3: Effect of combined data augmentation techniques, including rotation, cropping and the pro￾posed Semantic Relevant Classes Enhancement. On the left, the original image and SSM and on the right the resulting augmented version. To effectively train the SQ-GAN , a multi-…
Figure 5.4
Figure 5.4. Figure 5.4: Schematic representation of the semantic generator network Gs training pipeline. The process begins by collecting all "traffic signs" and "traffic lights" from the images and SSMs in the current mini-batch. For each image in the mini-batch, a random number n between …
Figure 5.5
Figure 5.5. Figure 5.5: Schematic representation of the image generator network Gx training pipeline. ranging from 5% to 100% with and expected value of 35%. This approach allows the model to learn to compress the SSM at various levels of compression. The sub-network Gs is trained using the…
Figure 5.6
Figure 5.6. Figure 5.6: Visual representation of the latent tensor selection of the SAMM projected in the image and SSM domain. In both cases the masking has been fixed to mx = ms = 0.20 and the region considered semantically relevant are shown on the right. Original mx = 0.95, ms = 0.15 mx…
Figure 5.7
Figure 5.7. Figure 5.7: Visual comparison between the same image and SSM at different masking reactions mx and ms. The original image and SSM are shown on the left. The upper row shows the reconstructed xˆ and the generated SSM using the SOTA SS-Model INTERN-2.5 [116]. The bottom row shows …
Figure 5.8
Figure 5.8. Figure 5.8: SSM retention evaluated between the true s and the reconstructed ˆs with the mIoU met￾ric as a function of the masking fraction ms. As the masking fraction ms increases the network Gs is able to better reconstruct the SSM . However, the increase of performances reach…
Figure 5.10
Figure 5.10. Figure 5.10: Visual comparison at different compression rates between the proposed SQ-GAN (TOP) and the classical BPG (BOTTOM). The SSMs shown are generated from xˆ via the SOTA SS-Model INTERN-2.5 [116]. The proposed model is able to reconstruct images with higher semantic rete…
Figure 5.11
Figure 5.11. Figure 5.11: Performance comparison between BPG, JPEG2000 and SQ-GAN in term of semantic metrics and the classic pixel-by-pixel PSNR. rithms by learning to select relevant parts of the image focusing on key elements and considering less relevant parts only when asked to. In summ…
Figure 6.1
Figure 6.1. Figure 6.1: EN scheme with connected EDs transmitting signals and receiving feedback from the ES. approaches are proposed: pure GOC via the GIB problem and SemGOC through the SQ-GAN model. The GIB framework is particularly well-suited for GOC , as detailed in Section 2.4.1. By i…
Figure 6.2
Figure 6.2. Figure 6.2: (a) The effects of Γk ∈ {0.5, 1, 8, 30, 50} on the power consumption of the ED and the ES. In this plot the value of G avg k changes as seen in the legend while the delay is fixed to D avg k = 0.003s. (b) The trade-off between device power consumption as a function o…
Figure 6.3
Figure 6.3. Figure 6.3: The trade-off between average power consumption, delay, and LPIPS metric. function in (6.18), the resulting problem can be optimized as a single-variable optimization problem. The final solution can be found by using a numerical approach. Once the optimal solutions f…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

127 extracted references · 70 canonical work pages

  1. [1]

    Agheli, N

    P. Agheli, N. Pappas, and M. Kountouris. Goal-oriented multiple access connectivity for networked intelligent systems. IEEE Communications Letters , PP:1--1, 08 2024

    2024

  2. [2]

    Akbari, J

    M. Akbari, J. Liang, and J. Han. Dsslic: Deep semantic segmentation-based layered image compression. In ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages 2042--2046, 2019

    2019

  3. [3]

    Aliyu, A

    A. Aliyu, A. Abdullah, O. Kaiwartya, Y. Cao, J. Mauri, N. Aslam, and M. Usman. Towards video streaming in iot environments: Vehicular communication perspective. Comput. Commun. , 118:93--119, 2017

    2017

  4. [4]

    S. B. B, P. Balaji, A. Munshi, W. Almukadi, T. N. Prabhu, V. K, and M. Abouhawwash. Machine learning based iot system for secure traffic management and accident detection in smart cities. PeerJ Computer Science , 9, 2023

    2023

  5. [5]

    Bahdanau, K

    D. Bahdanau, K. Cho, and Y. Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473 , 2014

    arXiv 2014

  6. [6]

    J. Bao, P. Basu, M. Dean, C. Partridge, A. Swami, W. Leland, and J. A. Hendler. Towards a theory of semantic communication. In 2011 IEEE Network Science Workshop , pages 110--117, 2011

    2011

  7. [7]

    F. Bellard. Better portable graphics image format. (http://bellard.org/bpg/), 2017

    2017

  8. [8]

    Binucci, P

    F. Binucci, P. Banelli, P. D. Lorenzo, and S. Barbarossa. Multi-user goal-oriented communications with energy-efficient edge resource management. IEEE Transactions on Green Communications and Networking , 7(4):1709--1724, 2023

    2023

Show all 127 references
  1. [9]

    Bourtsoulatze, D

    E. Bourtsoulatze, D. B. Kurka, and D. Gündüz. Deep joint source-channel coding for wireless image transmission. In ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages 4774--4778, 2019

    2019

  2. [10]

    Bruna, P

    J. Bruna, P. Sprechmann, and Y. LeCun. Super-resolution with deep convolutional sufficient statistics. In 4th International Conference on Learning Representations, ICLR 2016 ; Conference date: 02-05-2016 Through 04-05-2016 , Jan. 2016

    2016

  3. [11]

    T. D. Burd and R. W. Brodersen. Processor design for portable systems. J. VLSI Signal Process. Syst. , 13(2-3):203--221, Dec. 1996

    1996

  4. [12]

    Carnap and Y

    R. Carnap and Y. Bar - Hillel. An outline of a theory of semantic information. Journal of Symbolic Logic , 19(3):230--232, 1954

    1954

  5. [13]

    Chechik, A

    G. Chechik, A. Globerson, N. Tishby, and Y. Weiss. Information bottleneck for gaussian variables. In S. Thrun, L. Saul, and B. Sch\" o lkopf, editors, Advances in Neural Information Processing Systems , volume 16. MIT Press, 2004

    2004

  6. [14]

    Y. Chen, G. Li, C. Jin, S. Liu, and T. Li. Ssd-gan: Measuring the realness in the spatial and spectral domains. In AAAI , 2021

    2021

  7. [15]

    T. S. Chis and P. Harrison. Performance-energy trade-offs in smartphones. Proceedings of the 19th ACM International Conference on Modeling, Analysis and Simulation of Wireless and Mobile Systems , 2016

    2016

  8. [16]

    Cordts, M

    M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Benenson, U. Franke, S. Roth, and B. Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2016

    2016

  9. [17]

    T. M. Cover and J. A. Thomas. Elements of Information Theory (Wiley Series in Telecommunications and Signal Processing) . Wiley-Interscience, USA, 2006

    2006

  10. [18]

    Dhariwal and A

    P. Dhariwal and A. Q. Nichol. Diffusion models beat GAN s on image synthesis. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. W. Vaughan, editors, Advances in Neural Information Processing Systems , 2021

    2021

  11. [19]

    C. Dong, C. C. Loy, K. He, and X. Tang. Image super-resolution using deep convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence , 38(2):295--307, 2016

    2016

  12. [20]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning R...

    2021

  13. [21]

    Dowson and B

    D. Dowson and B. Landau. The fréchet distance between multivariate normal distributions. Journal of Multivariate Analysis , 12(3):450--455, 1982

    1982

  14. [22]

    Esser, R

    P. Esser, R. Rombach, and B. Ommer. Taming transformers for high-resolution image synthesis. In CVPR , pages 12873--12883. Computer Vision Foundation / IEEE, 2021

    2021

  15. [23]

    Felix, S

    A. Felix, S. Cammerer, S. Dörner, J. Hoydis, and S. Ten Brink. Ofdm-autoencoder for end-to-end learning of communications systems. In 2018 IEEE 19th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC) , pages 1--5, 2018

    2018

  16. [24]

    L. A. Gatys, A. S. Ecker, and M. Bethge. Texture synthesis using convolutional neural networks. In Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 1 , NIPS'15, page 262–270, Cambridge, MA, USA, 2015. MIT Press

    2015

  17. [25]

    Goodfellow, J

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial nets. In Advances in neural information processing systems , pages 2672--2680, 2014

    2014

  18. [26]

    Grigorescu, B

    S. Grigorescu, B. Trasnea, T. Cocias, and G. Macesanu. A survey of deep learning techniques for autonomous driving. Journal of Field Robotics , 37(3):362--386, 2020

    2020

  19. [27]

    Gündüz, Z

    D. Gündüz, Z. Qin, I. E. Aguerri, H. S. Dhillon, Z. Yang, A. Yener, K. K. Wong, and C.-B. Chae. Beyond transmitting bits: Context, semantics, and task-oriented communications. IEEE Journal on Selected Areas in Communications , 41(1):5--41, 2023

    2023

  20. [28]

    T. Han, Q. Yang, Z. Shi, S. He, and Z. Zhang. Semantic-preserved communication system for highly efficient speech transmission. IEEE Journal on Selected Areas in Communications , 41:245--259, 2022

    2022

  21. [29]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 770--778, 2016

    2016

  22. [30]

    Q. He, H. Yuan, D. Feng, B. Che, Z. Chen, and X. Xia. Robust semantic transmission of images with generative adversarial networks. GLOBECOM 2022 - 2022 IEEE Global Communications Conference , pages 3953--3958, 2022

    2022

  23. [31]

    Heusel, H

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems , volume 30. Curran Associates, Inc., 2017

    2017

  24. [32]

    Higgins, L

    I. Higgins, L. Matthey, A. Pal, C. P. Burgess, X. Glorot, M. M. Botvinick, S. Mohamed, and A. Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. In International Conference on Learning Representations , 2016

    2016

  25. [33]

    G. E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and R. R. Salakhutdinov. Improving neural networks by preventing co-adaptation of feature detectors. CoRR , abs/1207.0580, 2012. cite arxiv:1207.0580

    2012 arXiv

  26. [34]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems , volume 33, pages 6840--6851. Curran Associates, Inc., 2020

    2020

  27. [35]

    Ho and T

    J. Ho and T. Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications , 2021

    2021

  28. [36]

    Z. Hu, C. You, T. Liu, D. Wen, Y. Hu, Y. Cui, Y. Gong, and K. Huang. Semantic communication meets edge intelligence: Semantic-relay-aided text transmissions. IEEE Internet of Things Journal , pages 1--1, 2024

    2024

  29. [37]

    Huang, X

    L. Huang, X. Feng, L. Zhang, L. Qian, and Y. Wu. Multi-server multi-user multi-task computation offloading for mobile edge computing networks. Sensors , 19(6), 2019

    2019

  30. [38]

    Huang, Z

    M. Huang, Z. Mao, Q. Wang, and Y. Zhang. Not all image regions matter: Masked vector quantization for autoregressive image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 2002--2011, June 2023

    2002

  31. [39]

    Ioffe and C

    S. Ioffe and C. Szegedy. Batch normalization: accelerating deep network training by reducing internal covariate shift. In Proceedings of the 32nd International Conference on International Conference on Machine Learning - Volume 37 , ICML'15, page 448–456. JMLR.org, 2015

    2015

  32. [40]

    Isola, J.-Y

    P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros. Image-to-image translation with conditional adversarial networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 5967--5976, 2017

    2017

  33. [41]

    M. A. Ja'afreh, M. Alowaidi, H. A. Osman, and A. E. Saddik. Multimodal systems, experiences, and communications: A review toward the tactile internet vision. pages 191--220, 2018

    2018

  34. [42]

    M. A. Jensen. A history of mimo wireless communications. In 2016 IEEE International Symposium on Antennas and Propagation (APSURSI) , pages 681--682, 2016

    2016

  35. [43]

    Johnson, A

    J. Johnson, A. Alahi, and L. Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. volume 9906, pages 694--711, 10 2016

    2016

  36. [44]

    Jöckel, M

    L. Jöckel, M. Kläs, and S. Martínez-Fernández. Safe traffic sign recognition through data augmentation for autonomous vehicles software. In 2019 IEEE 19th International Conference on Software Quality, Reliability and Security Companion (QRS-C) , pages 540--541, 2019

    2019

  37. [45]

    Karras, S

    T. Karras, S. Laine, and T. Aila. A style-based generator architecture for generative adversarial networks. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 4396--4405, 2019

    2019

  38. [46]

    Kingma and J

    D. Kingma and J. Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR) , San Diega, CA, USA, 2015

    2015

  39. [47]

    D. P. Kingma and M. Welling. Auto-encoding variational bayes. In 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings , 2014

    2014

  40. [48]

    D. P. Kingma and M. Welling. An introduction to variational autoencoders. Found. Trends Mach. Learn. , 12(4):307--392, nov 2019

    2019

  41. [49]

    Konushin, B

    A. Konushin, B. Faizov, and V. Shakhuro. Road images augmentation with synthetic traffic signs using neural networks. Computer Optics , 5:736--748, 09 2021

    2021

  42. [50]

    Krizhevsky

    A. Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009

    2009

  43. [51]

    Krizhevsky, I

    A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In Proceedings of the 25th International Conference on Neural Information Processing Systems - Volume 1 , NIPS'12, page 1097–1105, Red Hook, NY, USA, 2012. Curran As...

    2012

  44. [52]

    LeCun, C

    Y. LeCun, C. Cortes, and C. Burges. Mnist handwritten digit database. ATT Labs [Online]. Available: http://yann.lecun.com/exdb/mnist , 2, 2010

    2010

  45. [53]

    LeCun, F

    Y. LeCun, F. J. Huang, and L. Bottou. Learning methods for generic object recognition with invariance to pose and lighting. Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition , 2:II--104 Vol.2, 2004

    2004

  46. [54]

    Ledig, L

    C. Ledig, L. Theis, F. Huszár, J. Caballero, A. Cunningham, A. Acosta, A. Aitken, A. Tejani, J. Totz, Z. Wang, and W. Shi. Photo-realistic single image super-resolution using a generative adversarial network. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (...

    2017

  47. [55]

    Image of a happy dog

    Lexthehead. Image of a happy dog. https://www.instagram.com/p/DBgsLletjvd/?img_index=2, 2024. Online image

    2024

  48. [56]

    B. Lim, S. Son, H. Kim, S. Nah, and K. M. Lee. Enhanced deep residual networks for single image super-resolution. In 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , pages 1132--1140, 2017

    2017

  49. [57]

    H. Liu, C. Li, Y. Li, and Y. J. Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 26296--26306, June 2024

    2024

  50. [58]

    S. Liu, L. Liu, J. Tang, B. Yu, Y. Wang, and W. Shi. Edge computing for autonomous driving: Opportunities and challenges. Proceedings of the IEEE , 107(8):1697--1716, 2019

    2019

  51. [59]

    K. Lu, Q. Zhou, R. Li, Z. Zhao, X. Chen, J. Wu, and H. Zhang. Rethinking modern communication from semantic coding to semantic communication. IEEE Wireless Communications , 30:158--164, 2021

    2021

  52. [60]

    Luo, H.-H

    X. Luo, H.-H. Chen, and Q. Guo. Semantic communications: Overview, open issues, and future research directions. IEEE Wireless Communications , 29(1):210--219, 2022

    2022

  53. [61]

    G. R. MacCartney, S. Deng, S. Sun, and T. S. Rappaport. Millimeter-wave human blockage at 73 ghz with a simple double knife-edge diffraction model and extension for directional antennas. In 2016 IEEE 84th Vehicular Technology Conference (VTC-Fall) , pages 1--6, 2016

    2016

  54. [62]

    Mach and Z

    P. Mach and Z. Becvar. Mobile edge computing: A survey on architecture and computation offloading. IEEE Communications Surveys and Tutorials , 19(3):1628--1656, 2017

    2017

  55. [63]

    Mercier, V

    H. Mercier, V. K. Bhargava, and V. Tarokh. A survey of error-correcting codes for channels with symbol synchronization errors. IEEE Communications Surveys and Tutorials , 12(1):87--96, 2010

    2010

  56. [64]

    Merluzzi, P

    M. Merluzzi, P. D. Lorenzo, and S. Barbarossa. Dynamic resource allocation for wireless edge machine learning with latency and accuracy guarantees. In ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages 9036--9040, 2020

    2020

  57. [65]

    Merluzzi, P

    M. Merluzzi, P. D. Lorenzo, and S. Barbarossa. Wireless edge machine learning: Resource allocation and trade-offs. IEEE Access , 9:45377--45398, 2021

    2021

  58. [66]

    Mirza and S

    M. Mirza and S. Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784 , 2014

    2014 arXiv

  59. [67]

    Mohammad and S

    U. Mohammad and S. Sorour. Adaptive task allocation for mobile edge learning. In 2019 IEEE Wireless Communications and Networking Conference Workshop (WCNCW) , pages 1--6, 2019

    2019

  60. [68]

    Mordacchini, M

    M. Mordacchini, M. Conti, A. Passarella, and R. Bruno. Human-centric data dissemination in the iop. ACM Transactions on Autonomous and Adaptive Systems (TAAS) , 14:1 -- 25, 2020

    2020

  61. [69]

    Nair and G

    V. Nair and G. E. Hinton. Rectified linear units improve restricted boltzmann machines. In ICML 2010 , pages 807--814, 2010

    2010

  62. [70]

    M. J. Neely. Stochastic Network Optimization with Application to Communication and Queueing Systems . Morgan and Claypool Publishers, 2010

    2010

  63. [71]

    A. Q. Nichol and P. Dhariwal. Improved denoising diffusion probabilistic models. In M. Meila and T. Zhang, editors, Proceedings of the 38th International Conference on Machine Learning , volume 139 of Proceedings of Machine Learning Research , pages 8162--8171. PMLR, 18--24 Jul 2021

    2021

  64. [72]

    Oluwasanmi, M

    A. Oluwasanmi, M. U. Aftab, A. Shokanbi, J. Jackson, B. Kumeda, and Z. Qin. Attentively conditioned generative adversarial network for semantic segmentation. IEEE Access , 8:31733--31741, 2020

    2020

  65. [73]

    Park, M.-Y

    T. Park, M.-Y. Liu, T.-C. Wang, and J.-Y. Zhu. Semantic image synthesis with spatially-adaptive normalization. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 2332--2341, 2019

    2019

  66. [74]

    Pascanu, T

    R. Pascanu, T. Mikolov, and Y. Bengio. On the difficulty of training recurrent neural networks. In Proceedings of the 30th International Conference on International Conference on Machine Learning - Volume 28 , ICML'13, pages III--1310--III--1318. JMLR.org, 2013

    2013

  67. [75]

    Pathak, P

    D. Pathak, P. Krähenbühl, J. Donahue, T. Darrell, and A. A. Efros. Context encoders: Feature learning by inpainting. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 2536--2544, 2016

    2016

  68. [76]

    E. Plaut. From principal subspaces to principal components with linear autoencoders. ArXiv , abs/1804.10253, 2018

    2018 arXiv

  69. [77]

    Podell, Z

    D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. Müller, J. Penna, and R. Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 07 2023

    2023

  70. [78]

    Z. Qin, X. Tao, J. Lu, and G. Y. Li. Semantic communications: Principles and challenges. ArXiv , abs/2201.01389, 2021

    2021 arXiv

  71. [79]

    Radford, J

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, and I. Sutskever. Language models are unsupervised multitask learners, 2019

    2019

  72. [80]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 10684--10695, 2022

    2022

  73. [81]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 10684--10695, June 2022

    2022

  74. [82]

    Ronneberger, P

    O. Ronneberger, P. Fischer, and T. Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation , pages 234--241. Springer International Publishing, Cham, 2015

    2015

  75. [83]

    D. E. Rumelhart, G. E. Hinton, and R. J. Williams. Learning internal representations by error propagation , page 318–362. MIT Press, Cambridge, MA, USA, 1986

    1986

  76. [84]

    Russakovsky, J

    O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual Recognition Challenge . International Journal of Computer Vision (IJCV) , 115(3):211--252, 2015

    2015

  77. [85]

    W. Saad, M. Bennis, and M. Chen. A vision of 6g wireless systems: Applications, trends, technologies, and open research problems. IEEE Network , 34(3):134--142, 2020

    2020

  78. [86]

    Salakhutdinov

    R. Salakhutdinov. Learning deep generative models. Annual Review of Statistics and Its Application , 2:361--385, 04 2015

    2015

  79. [87]

    Sampieri, G

    A. Sampieri, G. M. D. di Melendugno, A. Avogaro, F. Cunico, F. Setti, G. Skenderi, M. Cristani, and F. Galasso. Pose forecasting in industrial human-robot collaboration. In Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceeding...

    2022

  80. [88]

    Sandler, A

    M. Sandler, A. G. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 4510--4520, 2018

    2018

  81. [89]

    Satyanarayanan

    M. Satyanarayanan. The emergence of edge computing. Computer , 50(1):30--39, 2017

    2017

  82. [90]

    C. E. Shannon. A mathematical theory of communication. The Bell System Technical Journal , 27:379--423, 1948

    1948

  83. [91]

    J. Shao, Y. Mao, and J. Zhang. Learning task-oriented communication for edge inference: An information bottleneck approach. IEEE Journal on Selected Areas in Communications , 40(1):197--211, 2022

    2022

  84. [92]

    Y. Shao, Q. Cao, and D. Gündüz. A theory of semantic communication. IEEE Transactions on Mobile Computing , pages 1--18, 2024

    2024

  85. [93]

    W. Shi, J. Cao, Q. Zhang, Y. Li, and L. Xu. Edge computing: Vision and challenges. IEEE Internet of Things Journal , 3(5):637--646, 2016

    2016

  86. [94]

    Simonyan and A

    K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. In 3rd International Conference on Learning Representations (ICLR 2015) , pages 1--14. Computational and Biological Learning Society, 2015

    2015

  87. [95]

    Skatchkovsky and O

    N. Skatchkovsky and O. Simeone. Optimizing pipelined computation and communication for latency-constrained edge learning. IEEE Communications Letters , 23(9):1542--1546, 2019

    2019

  88. [96]

    Sneyers and P

    J. Sneyers and P. Wuille. Flif: Free lossless image format based on maniac compression. In 2016 IEEE International Conference on Image Processing (ICIP) , pages 66--70, 2016

    2016

  89. [97]

    Sohl-Dickstein, E

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In F. Bach and D. Blei, editors, Proceedings of the 32nd International Conference on Machine Learning , volume 37 of Proceedings of Machine Learning...

    2015

  90. [98]

    J. Song, C. Meng, and S. Ermon. Denoising diffusion implicit models. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021 . OpenReview.net, 2021

    2021

  91. [99]

    Stallkamp, M

    J. Stallkamp, M. Schlipsing, J. Salmen, and C. Igel. The G erman T raffic S ign R ecognition B enchmark: A multi-class classification competition. In IEEE International Joint Conference on Neural Networks , pages 1453--1460, 2011

    2011

  92. [100]

    Stockman and L

    G. Stockman and L. G. Shapiro. Computer Vision . Prentice Hall PTR, Upper Saddle River, NJ, USA, 2001

    2001

  93. [101]

    E. C. Strinati and S. Barbarossa. 6g networks: Beyond shannon towards semantic and goal-oriented communications. Computer Networks , 190:107930, 2021

    2021

  94. [102]

    E. C. Strinati, P. Di Lorenzo, V. Sciancalepore, A. Aijaz, M. Kountouris, D. Gündüz, P. Popovski, M. Sana, P. A. Stavrou, B. Soret, N. Cordeschi, S. Scardapane, M. Merluzzi, L. Zanzi, M. B. Renato, T. Quek, N. D. Pietro, O. Forceville, F. Costanzo, and P. Li. Goal-oriented and...

    2024

  95. [103]

    Szegedy, V

    C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna. Rethinking the inception architecture for computer vision. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 2818--2826, 2015

    2016

  96. [104]

    Tassi, M

    A. Tassi, M. Egan, R. Piechocki, and A. Nix. Modeling and design of millimeter-wave networks for highway vehicular communication. IEEE Transactions on Vehicular Technology , 66:10676--10691, 2017

    2017

  97. [105]

    Testa, C

    L. Testa, C. Battiloro, S. Sardellitti, and S. Barbarossa. Stability of graph convolutional neural networks through the lens of small perturbation analysis. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages 6865--6869, 2024

    2024

  98. [106]

    C. Tian, A. Steiner, S. Shamai, and S. N. Diggavi. Successive refinement via broadcast: Optimizing expected distortion of a gaussian source over a gaussian fading channel. IEEE Transactions on Information Theory , 54(7):2903--2918, 2008

    2008

  99. [107]

    Tishby, F

    N. Tishby, F. C. Pereira, and W. Bialek. The information bottleneck method. In Proc. of the 37-th Annual Allerton Conference on Communication, Control and Computing , pages 368--377, 1999

    1999

  100. [108]

    Tishby and N

    N. Tishby and N. Zaslavsky. Deep learning and the information bottleneck principle. In 2015 IEEE Information Theory Workshop (ITW) , pages 1--5, 2015

    2015

  101. [109]

    van den Oord, O

    A. van den Oord, O. Vinyals, and K. Kavukcuoglu. Neural discrete representation learning. In Proceedings of the 31st International Conference on Neural Information Processing Systems , NIPS'17, page 6309–6318, Red Hook, NY, USA, 2017. Curran Associates Inc

    2017

  102. [110]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems , NIPS'17, page 6000–6010, Red Hook, NY, USA, 2017. Cur...

    2017

  103. [111]

    Vincent, H

    P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, and P.-A. Manzagol. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. J. Mach. Learn. Res. , 11:3371--3408, Dec. 2010

    2010

  104. [112]

    J. Wang, W. bing Wu, Z. Liao, A. K. Sangaiah, and R. Sherratt. An energy-efficient off-loading scheme for low latency in collaborative edge computing. IEEE Access , 7:149182--149190, 2019

    2019

  105. [113]

    S. Wang, T. Tuor, T. Salonidis, K. K. Leung, C. Makaya, T. He, and K. Chan. When edge meets learning: Adaptive control for resource-constrained distributed machine learning. In IEEE INFOCOM 2018 - IEEE Conference on Computer Communications , pages 63--71, 2018

    2018

  106. [114]

    Wang, M.-Y

    T.-C. Wang, M.-Y. Liu, J.-Y. Zhu, A. Tao, J. Kautz, and B. Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2018

    2018

  107. [115]

    W. Wang, J. Bao, W. Zhou, D. Chen, D. Chen, L. Yuan, and H. Li. Semantic image synthesis via diffusion models. arXiv:2207.00050, 2022

    2022

  108. [116]

    W. Wang, J. Dai, Z. Chen, Z. Huang, Z. Li, X. Zhu, X. Hu, T. Lu, L. Lu, H. Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. arXiv preprint arXiv:2211.05778 , 2022

    2022 arXiv

  109. [117]

    W. Weaver. Recent contributions to the mathematical theory of communication. ETC: A Review of General Semantics , 10:261--281, 1953

    1953

  110. [118]

    J. Wu, C. Zhang, T. Xue, W. T. Freeman, and J. B. Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. In Proceedings of the 30th International Conference on Neural Information Processing Systems , NIPS'16, page 82–90, Red H...

    2016

  111. [119]

    H. Xie, Z. Qin, G. Y. Li, and B.-H. Juang. Deep learning enabled semantic communication systems. IEEE Transactions on Signal Processing , 69:2663--2675, 2021

    2021

  112. [120]

    B. Xu, N. Wang, T. Chen, and M. Li. Empirical evaluation of rectified activations in convolutional network. CoRR , abs/1505.00853, 2015

    2015 arXiv

  113. [121]

    L. Yang, J. Kang, X. Peng, B. Jia, Y. Mu, and Y. Jiang. Mitigation of multi-user interference in mimo-ofdm based visible light communication systems. In 2022 14th International Conference on Communication Software and Networks (ICCSN) , pages 69--74, 2022

    2022

  114. [122]

    Zaidi, I

    A. Zaidi, I. Estella-Aguerri, and S. Shamai (Shitz). On the information bottleneck problems: Models, connections, applications and information theoretic views. Entropy , 22(2), 2020

    2020

  115. [123]

    M. D. Zeiler and R. Fergus. Visualizing and understanding convolutional networks. In D. Fleet, T. Pajdla, B. Schiele, and T. Tuytelaars, editors, Computer Vision -- ECCV 2014 , pages 818--833, Cham, 2014. Springer International Publishing

    2014

  116. [124]

    Zhang, H

    C. Zhang, H. Zou, S. Lasaulce, W. Saad, M. Kountouris, and M. Bennis. Goal-oriented communications for the iot and application to data compression. IEEE Internet of Things Magazine , 5, 12 2022

    2022

  117. [125]

    Zhang, P

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR , 2018

    2018

  118. [126]

    Y. Zhou, Z. Wang, C. Fang, T. Bui, and T. L. Berg. Visual to sound: Generating natural sound for videos in the wild. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 3550--3558, 2018

    2018

  119. [127]

    Z. Zhou, X. Chen, E. Li, L. Zeng, K. Luo, and J. Zhang. Edge intelligence: Paving the last mile of artificial intelligence with edge computing. Proceedings of the IEEE , 107(8):1738--1762, 2019

    2019

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.