This paper presents a pruning method, Reconstruction Error Aware Pruning (REAP), to reduce the redundancy of convolutional neural network models for accelerating their inference. In REAP, we have the following steps: 1) Prune the channels whose outputs are redundant and can be reconstructed from the outputs of other channels in each convolutional layer; 2) Update the weights of the remaining channels by least squares method so as to compensate the error caused by pruning. This is how we compress and accelerate the models that are initially large and slow with little degradation. The ability of REAP to maintain the model performances saves us lots of time and labors for retraining the pruned models. The challenge of REAP is the computational cost for selecting the channels to be pruned. For selecting the channels, we need to solve a huge number of least squares problems. We have developed an efficient algorithm based on biorthogonal system to obtain the solutions of those least squares problems. In the experiments, we show that REAP can conduct pruning with smaller sacrifice of the model performances than several existing methods including the previously state-of-the-art one.

This paper presents a method for reducing the redundancy in both fully connected layers and convolutional layers of trained neural network models. The proposed method consists of two steps, 1) Neuro-Coding: to encode the behavior of each neuron by a vector composed of its outputs corresponding to actual inputs and 2) Neuro-Unification: to unify the neurons having the similar behavioral vectors. Instead of just pruning one of the similar neurons, the proposed method let the remaining neuron emulate the behavior of the pruned one. Therefore, the proposed method can reduce the number of neurons with small sacrifice of accuracy without retraining. Our method can be applied for compressing convolutional layers as well. In the convolutional layers, the behavior of each channel is encoded by its output feature maps, and channels whose behaviors can be well emulated by other channels are pruned and update the remaining weights. Through several experiments, we comfirmed that the proposed method performs better than the existing methods.

Pruning is an effective technique to reduce computational complexity of Convolutional Neural Networks (CNNs) by removing redundant neurons (or weights). There are two types of pruning methods: holistic pruning and layer-wise pruning. The former selects the least important neuron from the entire model and prunes it. The latter conducts pruning layer by layer. Recently, it has turned out that some layer-wise methods are effective for reducing computational complexity of pruned models while preserving their accuracy. The difficulty of layer-wise pruning is how to adjust pruning ratio (the ratio of neurons to be pruned) in each layer. Because CNNs typically have lots of layers composed of lots of neurons, it is inefficient to tune pruning ratios by human hands. In this paper, we present Pruning Ratio Optimizer (PRO), a method that can be combined with layer-wise pruning methods for optimizing pruning ratios. The idea of PRO is to adjust pruning ratios based on how much pruning in each layer has an impact on the outputs in the final layer. In the experiments, we could verify the effectiveness of PRO.