Concrete ML
WebsiteLibrariesProducts & ServicesDevelopersSupport
1.5
1.5
  • Welcome
  • Getting Started
    • What is Concrete ML?
    • Installation
    • Key concepts
    • Inference in the cloud
  • Built-in Models
    • Linear models
    • Tree-based models
    • Neural networks
    • Nearest neighbors
    • Encrypted dataframe
    • Encrypted training
  • Deep Learning
    • Using Torch
    • Using ONNX
    • Step-by-step guide
    • Debugging models
    • Optimizing inference
  • Guides
    • Prediction with FHE
    • Production deployment
    • Hybrid models
    • Serialization
  • Tutorials
    • See all tutorials
    • Built-in model examples
    • Deep learning examples
  • References
    • API
    • Pandas support
  • Explanations
    • Quantization
    • Pruning
    • Compilation
    • Advanced features
    • Project architecture
      • Importing ONNX
      • Quantization tools
      • FHE Op-graph design
      • External libraries
  • Developers
    • Set up the project
    • Set up Docker
    • Documentation
    • Support and issues
    • Contributing
    • Support new ONNX node
    • Release note
    • Feature request
    • Bug report
Powered by GitBook

Libraries

  • TFHE-rs
  • Concrete
  • Concrete ML
  • fhEVM

Developers

  • Blog
  • Documentation
  • Github
  • FHE resources

Company

  • About
  • Introduction to FHE
  • Media
  • Careers
On this page
  • Circuit bit-width optimization
  • Structured pruning
  • Rounded activations and quantizers
  • TLU error tolerance adjustment

Was this helpful?

Export as PDF
  1. Deep Learning

Optimizing inference

PreviousDebugging modelsNextPrediction with FHE

Last updated 1 year ago

Was this helpful?

Neural networks pose unique challenges with regards to encrypted inference. Each neuron in a network applies an activation function that requires a PBS operation. The latency of a single PBS depends on the bit-width of the input of the PBS.

Several approaches can be used to reduce the overall latency of a neural network.

Circuit bit-width optimization

and introduce specific hyper-parameters that influence the accumulator sizes. It is possible to chose quantization and pruning configurations that reduce the accumulator size. A trade-off between latency and accuracy can be obtained by varying these hyper-parameters as described in the .

Structured pruning

While un-structured pruning is used to ensure the accumulator bit-width stays low, can eliminate entire neurons from the network. Many neural networks are over-parametrized (since this enables easier training) and some neurons can be removed. Structured pruning, applied to a trained network as a fine-tuning step, can be applied to built-in neural networks using the helper function as shown in . To apply structured pruning to custom models, it is recommended to use the package.

Rounded activations and quantizers

Reducing the bit-width of the inputs to the Table Lookup (TLU) operations is a major source of improvements in the latency. Post-training, it is possible to leverage some properties of the fused activation and quantization functions expressed in the TLUs to further reduce the accumulator. This is achieved through the rounded PBS feature as described in the . Adjusting the rounding amount, relative to the initial accumulator size, can bring large improvements in latency while maintaining accuracy.

TLU error tolerance adjustment

Finally, the TFHE scheme exposes a TLU error tolerance parameter that has an impact on crypto-system parameters that influence latency. A higher tolerance of TLU off-by-one errors results in faster computations but may reduce accuracy. One can think of the error of obtaining T[x]T[x]T[x] as a Gaussian distribution centered on xxx: TLU[x]TLU[x]TLU[x] is obtained with probability of 1 - p_error, while T[x−1]T[x-1]T[x−1], T[x+1]T[x+1]T[x+1] are obtained with much lower probability, etc. In Deep NNs, these type of errors can be tolerated up to some point. See the and more specifically the usage example of .

Quantization Aware Training
pruning
structured pruning
prune
this example
torch-pruning
rounded activations and quantizers reference
p_error documentation for details
the API for finding the best p_error
deep learning design guide