# Nota’s model-compression patent: retraining a smaller layer group with cached inputs and outputs

A technical reading of US12361264B2: selecting layers, retaining intermediate features and retraining locally, with the claimed LLM configuration distinguished from an illustrative inspection camera.

Source: https://www.iplexlaw.co.kr/en/blog/nota-local-model-compression-cached-features

HOME / NEWS & INSIGHTS NEWS & INSIGHTS Nota’s model-compression patent: retraining a smaller layer group with cached inputs and outputs A technical reading of US12361264B2: selecting layers, retaining intermediate features and retraining locally, with the claimed LLM configuration distinguished from an illustrative inspection camera. Artificial Intelligence 2026.11.09 published IPLEX 7 min read Conceptual illustration of reducing a selected layer group and readjusting it with its original input–output pairs. Deploying an AI model on an edge device may require reducing its memory and processing demands. Nota’s US12361264B2 describes selecting part of a model, storing its original inputs and outputs, and using those records to retrain the compressed part. A hypothetical inspection camera helps explain what is reduced, what guides the adjustment and which results should be compared before deployment. 1. Moving an inspection model from a server to an edge device Imagine a factory camera checking packages as boxes pass along a conveyor. A model that runs comfortably on a development server may occupy too much memory or respond too slowly on the small device beside the production line. It must process each image and deliver a decision before the next box arrives, so performance on the actual device matters. Model compression reduces computations or stored values to make a trained model easier to run with limited resources. Changing one part of the model can change the values passed to later layers and, ultimately, its decisions. The resulting engineering question is how to recover information after reducing an expensive computation. The patent describes retaining the inputs and outputs of a selected layer group, then using them to retrain that group after compression. 2. Selecting a layer group for local compression A neural-network layer receives numerical inputs, performs a computation and passes its output onward. The blocks numbered 1, 2 and 3 in the illustration represent connected stages. A layer group defines the part treated together during compression and retraining; one illustrated block may contain several actual layers. The patent discusses a large language model comprising transformer blocks. A residual connection carries a block’s input forward, and the block’s computed result is added to it. The described selection procedure assesses relative importance by comparing the residual input with the sum of that input and the block’s result. Cosine similarity is one disclosed measure: thinking of two sets of numbers as arrows, it measures how closely their directions align. Higher similarity is assigned lower relative importance in this example. Choosing the candidate group, compressing it and adjusting its output are distinct stages. 3. Capturing the original group’s input–output pairs Example data first pass through the pretrained model. The system stores the values entering the first layer of the selected group and the values leaving its last layer for the same example. Each record therefore links an original input to its corresponding output. These values are usually intermediate features rather than an unchanged photograph or sentence. Earlier computations may encode colour changes, shapes or aspects of linguistic context as numerical representations needed by the next stage. The retained features are those actually crossing the selected group’s boundaries. The two cards in the conceptual diagram represent this pair. The input card supplies data to the smaller group; the output card records what the original group produced. Together they establish an input and a target for retraining. 4. Retraining the compressed group against the stored output The disclosure gives pruning, filter decomposition and layer removal as examples of compression. If removal would eliminate every layer in the selected group, preceding or following layers can be included so that there are still layers to retrain. Defining the group makes the later training scope explicit. Feeding the stored input features into the smaller group produces a new output. Retraining adjusts the group’s weights to bring that output closer to the stored output of the original group. The disclosure allows the stored data themselves, or a portion of them, to serve as the retraining inputs and targets. The adjusted group occupies the original group’s position in the model. Earlier layers provide its input features, and later layers receive its output. Here, “local” refers to concentrating compression and retraining on a defined part of the model. 5. A hypothetical camera example makes the intermediate comparison visible Simplify package inspection into three stages: extracting edges and surface changes, representing the relationship between tape and box surfaces, and making the inspection decision. Suppose the development team wants to reduce the computations in the middle stage. It first runs photographs of normal boxes and boxes with folded tape through the original model, retaining pairs of features entering and leaving that middle stage. After shrinking the middle block, it feeds the same input features into it and trains toward the original block’s recorded outputs. Hypothetical inspection-camera example. The dotted outline indicates original output features used as the target when retraining the compressed part. The top of the illustration represents the original path and the bottom the compressed path. The dotted output denotes the retraining target. If a feature relevant to folded tape is lost, comparing the new output with the stored target helps identify the change. The complete model must also be run to compare its actual inspection decisions. 6. NetsPresso’s development, optimization and device-testing workflow Nota’s official NetsPresso description separates model development, hardware-oriented optimization and testing on real devices. Its development-stage Compressor offers model compression such as structured pruning. The optimization stage includes graph optimization, quantization and conversion of model formats. The product materials also describe tools for evaluating performance on actual hardware. Devices differ in supported operations and memory arrangements. Reviewing model size and accuracy, then latency and executability on the target device, helps connect compression results to deployment decisions. In the camera example, a team could select the computations to reduce and compare the original and compressed models on the same device with the same image set. Cached intermediate features explain the patent’s retraining reference, while the product workflow illustrates how a prepared model may be assessed in its intended environment. 7. Documenting what changed and how the comparison was made Record the location of the compressed group and why it was selected: which layers remain, which computations changed, which situations the inputs cover and which stored outputs serve as targets. Separating normal examples from rare defects also makes subsequent analysis clearer. Closeness to the original intermediate output and preservation of final task accuracy are different measurements. The first describes local retraining; the second concerns the complete model’s task. Recording accuracy, processing time and memory use under consistent conditions makes both gains and remaining trade-offs visible. A useful technical account connects selection criteria, retained inputs and outputs, retraining scope and deployment comparisons. Explaining which part was changed to address which problem is more informative than stating only that the model became smaller. The original figure labels and forms are preserved; the text explains the source material. The camera is an explanatory example, not a reported implementation or benchmark. Independent claims 1 and 7 include an LLM with at least one billion parameters and transformer-block selection based on the specified similarity; cosine similarity is specified in dependent claims 4 and 10. The product description does not establish that NetsPresso implements every claimed element. USPTO patent text · US12361264B2 Nota’s official NetsPresso product overview DEEPX DX-M1 and edge AI: connecting training with deployment feedback Understanding LoRA: adapting an AI model with small weight updates Explore IPLEX’s AI practice and consultation materials to document model modifications and their technical significance. Explore our AI and patent practice Original Tistory article Read the Korean source This article reflects the information available when it was published. Contact us to discuss your circumstances. Discuss this topic ↗ All articles ON THIS PAGE 1. Moving an inspection model from a server to an edge device 2. Selecting a layer group for local compression 3. Capturing the original group’s input–output pairs 4. Retraining the compressed group against the stored output 5. A hypothetical camera example makes the intermediate comparison visible 6. NetsPresso’s development, optimization and device-testing workflow 7. Documenting what changed and how the comparison was made TALK TO IPLEX Discuss your IP questions We consider your technology and business needs together. ↗ Contact us Newer Overseas online offers and infringement of Korean patents ↗ Older Design licences and a subsequent transfer of ownership ↗ Related insights Artificial Intelligence 2026.11.19 Matryoshka embeddings for image retrieval: shortlist with a prefix, rerank with the full vector A practical reading of Matryoshka Representation Learning, its nested training objective and the ImageNet-1K retrieval results reported at NeurIPS 2022. ↗ Read article Artificial Intelligence 2026.11.18 ByteDance’s captioning patent: transcribing separate tracks and resolving overlapping display times How US11837234B2 processes selected audio tracks and adjusts caption timing, with a practical distinction between patent disclosure and CapCut’s public guidance. ↗ Read article Artificial Intelligence 2026.11.17 Apple recognized text: connect a phone number or web address to the right action US12001642B2 links recognized text, image coordinates, text types and user input to contextual menus and actions. ↗ Read article

- https://www.iplexlaw.co.kr/en
- https://www.iplexlaw.co.kr/en/blog/category/ai
- https://image-ppubs.uspto.gov/dirsearch-public/print/downloadPdf/12361264
- https://www.nota.ai/netspresso
- https://www.iplexlaw.co.kr/en/blog/deepx-heterogeneous-processors-edge-ai
- https://www.iplexlaw.co.kr/en/blog/lora-low-rank-adaptation
- https://www.iplexlaw.co.kr/en/ai-patents
- https://iplexlaw.tistory.com/20
- https://www.iplexlaw.co.kr/ko/blog/nota-local-model-compression-cached-features
- https://www.iplexlaw.co.kr/en/contact
- https://www.iplexlaw.co.kr/en/blog
- https://www.iplexlaw.co.kr/en/blog/nota-local-model-compression-cached-features#article-20-en-0
- https://www.iplexlaw.co.kr/en/blog/nota-local-model-compression-cached-features#article-20-en-1
- https://www.iplexlaw.co.kr/en/blog/nota-local-model-compression-cached-features#article-20-en-2
- https://www.iplexlaw.co.kr/en/blog/nota-local-model-compression-cached-features#article-20-en-3
- https://www.iplexlaw.co.kr/en/blog/nota-local-model-compression-cached-features#article-20-en-4
- https://www.iplexlaw.co.kr/en/blog/nota-local-model-compression-cached-features#article-20-en-5
- https://www.iplexlaw.co.kr/en/blog/nota-local-model-compression-cached-features#article-20-en-6
- https://www.iplexlaw.co.kr/en/contact
- https://www.iplexlaw.co.kr/en/blog/brunch-case-1936
- https://www.iplexlaw.co.kr/en/blog/brunch-case-1935
- https://www.iplexlaw.co.kr/en/blog/matryoshka-prefix-vector-image-retrieval
- https://www.iplexlaw.co.kr/en/blog/matryoshka-prefix-vector-image-retrieval
- https://www.iplexlaw.co.kr/en/blog/bytedance-track-based-caption-timing
- https://www.iplexlaw.co.kr/en/blog/bytedance-track-based-caption-timing
- https://www.iplexlaw.co.kr/en/blog/apple-recognized-text-contextual-actions
- https://www.iplexlaw.co.kr/en/blog/apple-recognized-text-contextual-actions
