Patent Drafting Analysis of WaveOne Inc.’s GAN-Based Compression Training System | US 2018/0174052 A1
Patent Drafting Analysis of WaveOne Inc.'s GAN-Based Compression Training System | US 2018/0174052 A1
A structural and strategic analysis of WaveOne's foundational GAN compression patent, examining claim architecture, drafting quality, §101 eligibility exposure, critical gaps, and prosecution positioning across method and CRM claim types.
Spec Words
6,200
Across 6 sections
Draft now ↗
Total Claims
25
3 independent · 22 dependent
Draft now ↗
Figure Sheets
7
System architecture, training pipelines, discriminator architecture, deployment flow
Draft now ↗
Published by PatSnap Insights Team · · 12 min read Verified by PatSnap Eureka Data
Overview
Structural Overview
The detailed description dominates at approximately 52% of total specification words (~3,200 of ~6,200), with the claims section contributing a substantial 34% (~2,130 words), reflecting the repetitive parallel claim structure across three independent claims. The patent includes 25 claims total — 3 independent and 22 dependent — spread across method (Claim 1), CRM-decoder (Claim 9), and CRM-encoder (Claim 17) types. Seven figure sheets cover training architecture, discriminator design, multi-scale embodiments, and deployment, providing reasonable but not exhaustive structural support.
Section Word Distribution
↗ Click bars to exploreFigure Inventory — 7 Sheets
| Figure | Description | Role |
|---|---|---|
| FIG. 1 | Block diagram of the system environment 100 showing the compression system 130, sender system 110, receiver system 140, and network 120 interconnection.Search in Eureka ↗ | System architecture |
| FIG. 2 | General training process diagram showing the autoencoder 202 coupled to discriminator 204, with reconstruction loss 230, discriminator loss 234, codelength regularization loss 238, and code outputs 282.Search in Eureka ↗ | Flow diagram |
| FIG. 3 | Detailed training process showing encoder 350 and decoder 360 within autoencoder 302, discriminator 304, adaptive coder 380, and the tensor y with CxHxW dimensions feeding into compressed code 382.Search in Eureka ↗ | Claim support |
| FIG. 4 | Architecture of the discriminator 304 showing multi-layer neural network (layers 1, 5, 10, 15, 20) combining intermediate outputs to produce the discrimination prediction.Search in Eureka ↗ | Key embodiment |
| FIG. 5 | Alternative detailed training process showing autoencoder 502 with discriminator 504 receiving ordered pairs via uniform swap mechanism 550, adaptive coder 580, and codelength regularization loss 538.Search in Eureka ↗ | Claim support |
| FIG. 6 | Deployment process for encoder 650 and decoder 660 showing sender system 110 applying encoder and entropy coder 680 to generate compressed code 682 transmitted over network to receiver system 140 with entropy decoder 690.Search in Eureka ↗ | System architecture |
| FIG. 7 | Multi-scale training process showing three autoencoder-discriminator pairs (702/704, 702'/704', 702"/704") operating at scales s1, s2, s3, with reconstructed content combined via summation nodes to produce final output 730.Search in Eureka ↗ | Key embodiment |
Analysis powered by PatSnap Eureka. Patent text and figures publicly available from USPTO. Draft a Similar Patent
Claims
Claim Architecture Analysis
The patent contains 3 independent claims: Claim 1 (method for training encoder and decoder), Claim 9 (decoder stored on computer readable storage medium manufactured by training process), and Claim 17 (encoder stored on computer readable storage medium manufactured by training process). The dependent-to-independent ratio of 7.33:1 is above the software/AI art unit norm, providing layered fallback coverage, though the 22 dependent claims are largely parallel repetitions across the three independent claim tracks rather than genuinely distinct fallback positions. The tripartite method/CRM-decoder/CRM-encoder structure provides enforcement coverage across training, deployment-decoder, and deployment-encoder scenarios but notably lacks a system or apparatus claim.
Core inventive concept: Claims 1, 9, and 17 each solve the problem of compression artifacts in reconstructed content by training an autoencoder (encoding portion + decoding portion) using a combined loss function that includes both a reconstruction loss — measuring dissimilarity between training content and reconstructed content — and a discriminator loss — measuring the cost of incorrect discrimination predictions generated by applying a GAN-style discriminator portion to input content comprising both the training content and the reconstructed content, with backpropagation stopped when the loss function satisfies a predetermined criteria.
Independent Claim Dissection
| Claim | Preamble | Transition | Key Body Elements |
|---|---|---|---|
| Claim 1 | A method for training an encoder and a decoder | comprising | obtaining a set of training content; compression model with encoding portion, decoding portion, and discriminator portion; repeatedly backpropagating error terms from loss function (reconstruction loss + discriminator loss) to update encoding and decoding portion parameters; stopping backpropagation after loss function satisfies predetermined criteriaSearch prior art ↗ |
| Claim 9 | A decoder stored on a computer readable storage medium, wherein the decoder is manufactured by a process | comprising | obtaining training content; compression model with encoding, decoding, and discriminator portions; backpropagating error terms from combined loss function to update encoding and decoding parameters; stopping at predetermined criteria; storing decoding portion parameters as decoder parameters; decoder coupled to receive compressed code and output reconstructed contentSearch prior art ↗ |
| Claim 17 | An encoder stored on a computer readable storage medium, wherein the encoder is manufactured by a process | comprising | obtaining training content; compression model with encoding, decoding, and discriminator portions; backpropagating error terms from combined loss function to update encoding and decoding parameters; stopping at predetermined criteria; storing encoding portion parameters as encoder parameters; encoder coupled to receive content and output compressed codeSearch prior art ↗ |
Claim Dependency Tree
1 Method for training encoder/decoder — compression model with encoding, decoding, discriminator portions; loss function includes reconstruction loss + discriminator lossSearch Claim 1 prior art ↗
2 Adds: discriminator receives each training/reconstructed content individually; predictions indicate likelihood content is reconstructed versionSearch in Eureka ↗
3 Adds: discriminator receives ordered pairs of training content and corresponding reconstructed content; predictions indicate which in ordered pair is reconstructedSearch in Eureka ↗
4 Further: ordered pairs include first pair with training content first / reconstructed second, and second pair with reconstructed first / training secondSearch in Eureka ↗
5 Adds: discriminator portion includes neural network model; predictions generated by combining outputs from intermediate layersSearch in Eureka ↗
6 Adds: loss function further includes codelength regularization loss based on magnitudes of tensor elementsSearch in Eureka ↗
7 Adds: separately backpropagating discriminator loss to update discriminator parameters while fixing encoder/decoder parametersSearch in Eureka ↗
8 Further: adaptive training based on discriminator accuracy — above first threshold trains encoder/decoder; below second threshold trains discriminator; between thresholds alternatesSearch in Eureka ↗
9 Decoder on computer readable storage medium manufactured by GAN-training process; decoder stores decoding portion parameters; receives compressed code, outputs reconstructed contentSearch Claim 9 prior art ↗
10 Adds: discriminator receives training/reconstructed content individually; predictions indicate likelihood content is reconstructedSearch in Eureka ↗
11 Adds: discriminator receives ordered pairs; predictions indicate which in pair is reconstructedSearch in Eureka ↗
12 Further: ordered pairs include first pair (training first, reconstructed second) and second pair (reconstructed first, training second)Search in Eureka ↗
13 Adds: discriminator includes neural network; predictions from combining intermediate layer outputsSearch in Eureka ↗
14 Adds: codelength regularization loss in loss function based on tensor element magnitudesSearch in Eureka ↗
15 Adds: separate discriminator backpropagation while fixing encoder/decoder parametersSearch in Eureka ↗
16 Further: adaptive threshold-based alternation between encoder/decoder training and discriminator trainingSearch in Eureka ↗
17 Encoder on computer readable storage medium manufactured by GAN-training process; encoder stores encoding portion parameters; receives content, outputs compressed codeSearch Claim 17 prior art ↗
18 Adds: discriminator receives training/reconstructed content individually; predictions indicate likelihood content is reconstructedSearch in Eureka ↗
19 Adds: discriminator receives ordered pairs; predictions indicate which in pair is reconstructedSearch in Eureka ↗
20 Further: ordered pairs with training first/reconstructed second and reconstructed first/training secondSearch in Eureka ↗
21 Adds: discriminator includes neural network; predictions from combining intermediate layer outputsSearch in Eureka ↗
22 Adds: codelength regularization loss in loss functionSearch in Eureka ↗
23 Adds: separate discriminator backpropagation while fixing encoder/decoder parametersSearch in Eureka ↗
24 Further: adaptive threshold-based alternation between encoder/decoder and discriminator trainingSearch in Eureka ↗
25 Method for training encoder/decoder — multi-scale: downsampling training content into scales; set of autoencoder-discriminator pairs; each pair's loss function includes reconstruction loss (from combined multi-scale output) and discriminator loss at corresponding scaleSearch Claim 25 prior art ↗
| Metric | This Application | Software / AI Industry Norm |
|---|---|---|
| Total claims | 25 | 15 – 25 |
| Independent claim count | 4 | 2 – 4 |
| Dependent : Independent ratio | 5.25 : 1 | 4 – 8 : 1 |
| Method claims present? | Yes — Claims 1, 25 | Common |
| System / apparatus claims? | No | Common |
Analysis powered by PatSnap Eureka. Patent text and figures publicly available from USPTO. Draft a Similar Patent
Drafting Quality
Drafting Quality Signals
The GAN compression claim set demonstrates strong structural clarity in Claim 1's loss function definition — the combined reconstruction loss and discriminator loss elements are precisely enumerated — but the absence of any apparatus or system claim creates a significant enforcement gap. The parallel triplication of dependent claims 2–8 across Claims 9 and 17 inflates the claim count without adding substantive fallback positions, while the omission of a CRM claim covering the trained model file itself represents a missed protection vector.
✅
Antecedent Basis
The claim language is largely clean with respect to antecedent basis. In Claim 1, "the training content" has proper antecedent basis from "a set of training content" in the preamble step, "the reconstructed content" is introduced via the reconstruction loss limitation, and "the discriminator portion" follows "a discriminator portion" in the compression model recitation. Claims 9 and 17 mirror this structure consistently. No orphaned "the [element]" references were identified across the 25 claims.
✅
Spec–Claim Consistency
Key claim limitations map directly to specific spec disclosures: the combined loss function (reconstruction loss + discriminator loss) in Claim 1 is supported by FIG. 2 (loss components 230, 234) and ¶[0037]–[0039]; the ordered pair discriminator variant in Claims 3–4 is supported by FIG. 5 and ¶[0061]–[0062]; the multi-layer intermediate output combination in Claims 5/13/21 is supported by FIG. 4 and ¶[0049]; and the codelength regularization loss in Claims 6/14/22 maps to FIG. 3 (element 238) and ¶[0041]–[0057]. The multi-scale variant in Claim 25 is fully supported by FIG. 7 and ¶[0070]–[0079].
✅
Transition Word Usage
All three independent claims (1, 9, 17, 25) use "comprising" as the transition, which is the strategically correct open-ended choice for AI/software claims — it ensures that implementations adding further components (e.g., additional loss terms, perceptual loss) do not escape the claim scope. The use of "comprising" in a method claim (Claim 1) appropriately preserves coverage over systems that add steps beyond the enumerated ones. No "consisting of" or "consisting essentially of" language was used anywhere, which is appropriate given the breadth goals of this filing.
⚠️
§112(f) Means-Plus-Function Risk
No explicit "means for" language appears in the claims, reducing direct §112(f) exposure. However, the functional language in Claims 9 and 17 — reciting a "decoder" and "encoder" each "stored on a computer readable storage medium" and defined entirely by their functional output — may invite §112 written description scrutiny as the claims define the article by its manufacturing process rather than structural features. An examiner may argue that a "decoder manufactured by a process" invokes product-by-process interpretation, potentially narrowing the claims to the specific training process disclosed rather than any functionally equivalent decoder, which could undermine the CRM claim coverage strategy.
⚠️
§101 Eligibility Risk
Claims 1 and 25 are method claims covering a training algorithm — backpropagation with a combined loss function — which is abstract under Alice/Mayo step one as a mathematical relationship or algorithm for organizing human activity (improving compression). The hardware tie-in is weak: Claims 1 and 25 do not recite any hardware processor, computer, or storage medium, making them vulnerable to §101 rejection as abstract methods. Claims 9 and 17 fare better by anchoring the claims to a "computer readable storage medium," but the training process steps themselves remain algorithmic. A stronger filing would have included at least one apparatus claim with a processor limitation or added a Berkheimer-style factual assertion about non-routine computer implementation.
⚠️
Dependent Claim Fallback Quality
The 22 dependent claims are structured as three parallel ladders of 7 claims each (Claims 2–8 on Claim 1, Claims 10–16 on Claim 9, Claims 18–24 on Claim 17), with each ladder containing the same limitations in sequence. This parallelism means that if Claim 1 is invalidated, Claims 9 and 17 — which independently recite the same training process — serve as true fallback, but the dependent claims 2–8 add no protection beyond what 10–16 or 18–24 already provide. Claims 7–8, 15–16, and 23–24 (the adaptive accuracy-threshold training alternation) add genuinely distinct fallback value; by contrast, Claims 2/10/18 (individual input mode) and 5/13/21 (intermediate layer outputs) are redundant across the claim set without adding scope differentiation.
⚠️
Abstract Quality
The abstract accurately describes the compression system's GAN-training architecture — encoder/decoder autoencoder with discriminator, reconstruction artifacts reduction, and the training competition mechanism — but omits the novel loss function formulation (the combination of reconstruction loss and negative discriminator loss) that is the actual inventive contribution captured in Claim 1. An examiner reading only the abstract would identify a GAN applied to compression but would not identify the specific loss function structure or the adaptive threshold-based training alternation (Claims 7–8), which are the most prosecution-defensible elements of the claim set. This increases the risk of a prior art search that misses the specific combination.
✅
Figure Support Quality
The seven figure sheets provide strong collective support for the structural claim limitations. FIG. 2 directly maps to the general autoencoder-discriminator training loop in Claim 1; FIG. 3 maps to the encoder/decoder decomposition and the codelength loss (Claims 6/14/22); FIG. 4 maps to the multi-layer intermediate output combination (Claims 5/13/21); FIG. 5 maps to the ordered-pair discriminator variant (Claims 3–4); FIG. 6 maps to the deployment scenario in Claims 9 and 17; and FIG. 7 maps directly to Claim 25's multi-scale training. The adaptive threshold training mechanism in Claims 8/16/24 (discriminator accuracy thresholds L and U) is described in ¶[0059] but lacks a dedicated figure, representing the only structural limitation without direct figure support.
Analysis powered by PatSnap Eureka. Patent text and figures publicly available from USPTO. Draft a Similar Patent
Scorecard
Strategic Intent Scorecard
Multi-dimensional assessment of this application's patent strategy quality, based on claim structure, specification depth, and prosecution positioning.
Claim Breadth
3.8
Prosecution Defensibility
3
Spec–Claim Consistency
4.2
Dependent Claim Coverage
2.5
Claim Type Diversity
2.5
Figure Support Quality
4
Key observation: The highest-scoring dimension is Spec–Claim Consistency (4.2/5.0): every independent claim limitation — combined loss function, tensor generation, backpropagation alternation — maps to a specific figure (FIGs. 2–7) and numbered paragraph (¶[0037]–[0079]), providing robust written description support that would be difficult for an examiner to challenge. The lowest-scoring dimension is Dependent Claim Coverage (2.5/5.0): the 22 dependent claims are structurally organized as three parallel ladders that replicate the same fallback positions (individual vs. ordered-pair discriminator, intermediate layers, codelength loss, adaptive thresholds) across Claims 1, 9, and 17, meaning a competitor can design around the entire dependent claim set by addressing a single ladder. Practitioners analyzing this portfolio should note that a continuation filing adding apparatus claims with specific hardware limitations and system claims covering the deployed compression pipeline would substantially strengthen the enforcement posture, particularly post-Alice.
See how your own draft compares — Open Eureka IP Drafting →
Critical Gaps
3 Critical Gaps in This Claim Set
A senior-attorney lens on the three highest-priority structural weaknesses — what each exposes in prosecution and litigation, and what a stronger filing would have done differently.
🔒
3 Critical Gaps in This Claim Set
See the full attorney-level analysis of what this application leaves unprotected — and how to draft it more defensively for your own filings.
No apparatus claim on compression system §101 risk in hardware-free method claims Deployed inference system unclaimed
Unlock Full Analysis — Free Frequently asked questions
US 2018/0174052 A1 — key questions answered
Still have questions? PatSnap Eureka can answer them from patent data instantly. Search in Eureka
Disclaimer: This analysis is generated by PatSnap Eureka AI based on publicly available patent data from the USPTO. It does not constitute legal advice and should not be relied upon as such. Patent data may be subject to change as prosecution progresses. Scores and assessments reflect automated analysis and may not capture all relevant legal or technical nuances. Always consult a qualified patent attorney for formal legal opinions on patentability, freedom to operate, or infringement.
Ask anything about this patent.
PatSnap Eureka searches patents and data to answer instantly.
PatSnap Eureka searches patents and data to answer instantly.
Powered by PatSnap Eureka
Link copied to clipboard