This document introduces several approaches to reduce the overall latency of a neural network.
Neural networks are challenging for encrypted inference. Each neuron in a network has to apply an activation function that requires a Programmable Bootstrapping(PBS) operation. The latency of a single PBS depends on the bit-width of its input.
Quantization Aware Training and pruning introduce specific hyper-parameters that influence the accumulator sizes. You can chose quantization and pruning configurations to reduce the accumulator size. To obtain a trade-off between latency and accuracy, you can manually set these hyper-parameters as described in the deep learning design guide.
While using unstructured pruning ensures the accumulator bit-width stays low, structured pruning can eliminate entire neurons from the network as many neural networks are over-parametrized for easier training. You can apply structured pruning to a trained network as a fine-tuning step. This example demonstrates how to apply structured pruning to built-in neural networks using the prune helper function. To apply structured pruning to custom models, it is recommended to use the torch-pruning package.
Reducing the bit-width of inputs to the Table Lookup (TLU) operations significantly improves latency. Post-training, you can leverage properties of the fused activation and quantization functions in the TLUs to further reduce the accumulator size. This is achieved through the rounded PBS feature as described in the rounded activations and quantizers reference. Adjusting the rounding amount relative to the initial accumulator size can improve latency while maintaining accuracy.
Finally, the TFHE scheme introduces 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. You can think of the error of obtaining as a Gaussian distribution centered on : is obtained with probability of 1 - p_error, while , are obtained with much lower probability, etc. In Deep NNs, these type of errors can be tolerated up to some point. See the p_error documentation for details and more specifically the API for finding the best p_error.