# US 2020/0401898 A1 · Optimizing machine learning model performance > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 2020/0401898 A1 - Document ID: US20200401898A1 - Status: Published application - Kind: A1 - Application number: 16/905,541 - Filed: 2020-06-18 - Published: 2020-12-24 - Earliest claimed: 2019-06-18 - Assignees: QUALCOMM Incorporated - Applicants: QUALCOMM Incorporated - Inventors: Meghal VARIA - Web page: https://patentgrep.com/patent/US20200401898A1 ## CPC - G06N 3/063: using electronic means - G06N 3/084: Backpropagation, e.g. using gradient descent - G06N 3/045: Combinations of networks - G06N 3/0464: Convolutional networks [CNN, ConvNet] - G06N 3/08: Learning methods - G06N 3/09: Supervised learning - G06N 5/01: Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound - G06N 5/046: Forward inferencing; Production systems - G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS ## Abstract Certain aspects of the present disclosure provide techniques for receiving data defining a neural network; analyzing the data to determine a depth-first cut point for a depth-first traversal portion of an overall network traversal; performing depth-first traversal for the depth-first portion of the overall network traversal; and performing layer-based traversal for a layer-based portion of the overall network traversal. ## Claims ### Claim 1 A method, comprising: receiving data defining a neural network; analyzing the data to determine a depth-first cut point for a depth-first traversal portion of an overall network traversal; performing depth-first traversal for the depth-first traversal portion of the overall network traversal; and performing layer-based traversal for a layer-based portion of the overall network traversal. ### Claim 2 (depends on 1) The method of claim 1, further comprising: storing weight data and output data for the depth-first traversal portion in a local memory during the depth-first portion of the overall network traversal. ### Claim 3 (depends on 2) The method of claim 2, further comprising: determining the depth-first cut point based at least in part on a sum of a weight data size and an output data size for the depth-first traversal portion being less than a capacity of the local memory. ### Claim 4 (depends on 3) The method of claim 3, wherein the local memory is a tightly-coupled memory (TCM) associated with a processor configured to perform the depth-first traversal and the layer-based traversal. ### Claim 5 (depends on 1) The method of claim 1, wherein performing layer-based traversal for the layer-based portion of the overall network traversal comprises: storing weight values and input values from a non-local memory in a local memory; performing processing of a first layer of the layer-based portion of the overall network traversal using the weight values and the input values in order to generate output values; storing the output values in the local memory; and performing processing of a second layer of the layer-based portion of the overall network traversal using the weight values and the output values. ### Claim 6 (depends on 3) The method of claim 3, wherein the depth-first cut point is determined by a compiler based on the capacity of the local memory. ### Claim 7 A processing system, comprising: a memory comprising computer-executable instructions; and a processor configured to execute the computer-executable instructions and cause the processing system to: receive data defining a neural network; analyze the data to determine a depth-first cut point for a depth-first traversal portion of an overall network traversal; perform depth-first traversal for the depth-first traversal portion of the overall network traversal; and perform layer-based traversal for a layer-based portion of the overall network traversal. ### Claim 8 (depends on 7) The processing system of claim 7, wherein the processor is further configured to: store weight data and output data for the depth-first traversal portion in a local memory during the depth-first portion of the overall network traversal. ### Claim 9 (depends on 8) The processing system of claim 8, wherein the processor is further configured to determine the depth-first cut point based at least in part on a sum of a weight data size and an output data size for the depth-first traversal portion being less than a capacity of the local memory. ### Claim 10 (depends on 9) The processing system of claim 9, wherein the local memory is a tightly-coupled memory (TCM) associated with a processor configured to perform the depth-first traversal and the layer-based traversal. ### Claim 11 (depends on 7) The processing system of claim 7, wherein in order to perform the layer-based traversal for the layer-based portion of the overall network traversal, the processor is further configured to: store weight values and input values from a non-local memory in a local memory; perform processing of a first layer of the layer-based portion of the overall network traversal using the weight values and the input values in order to generate output values; store the output values in the local memory; and perform processing of a second layer of the layer-based portion of the overall network traversal using the weight values and the output values. ### Claim 12 (depends on 9) The processing system of claim 9, wherein the depth-first cut point is determined by a compiler based on the capacity of the local memory. ### Claim 13 A method, comprising: receiving data defining a neural network; receiving hardware characteristic data defining a plurality of hardware resources, including a local memory capacity; analyzing the data to determine a cut point for a depth-first traversal portion of an overall network traversal; determining a first cut point based at least in part on the data and the local memory capacity, wherein the first cut point is at a maximum depth of the neural network given the local memory capacity; determining a second cut point, wherein the second cut point is shallower than the first cut point, and wherein the second cut point increases hardware parallelism compared to the first cut point; and performing depth-first traversal for the depth-first portion of the overall network traversal using two or more hardware resources of the plurality of hardware resource and the second cut point. ### Claim 14 (depends on 13) The method of claim 13, further comprising: storing weight data and output data for the depth-first traversal portion in a local memory during the depth-first portion of the overall network traversal. ### Claim 15 (depends on 14) The method of claim 14, wherein the local memory is a tightly-coupled memory (TCM). ### Claim 16 (depends on 13) The method of claim 13, wherein performing depth-first traversal for the depth-first traversal portion of the overall network traversal does not require writing data to a non-local memory. ### Claim 17 (depends on 13) The method of claim 13, wherein the second cut point enables cone parallelization during the depth-first traversal for the depth-first portion of the overall network traversal. ### Claim 18 A method, comprising: performing, during a first time interval, a computational phase of a first inference on a first hardware resource; performing, during a second time interval, a transport phase of the first inference on a second hardware resource; performing, during the second time interval, a computational phase of a second inference on the first hardware resource; and performing, during a third time interval, a transport phase of the second inference on the second hardware resource. ### Claim 19 (depends on 18) The method of claim 18, wherein the first hardware resource is optimized for inference computation. ### Claim 20 (depends on 19) The method of claim 19, wherein the second hardware resource is optimized for data transport.