Computationally Efficient Measures of Internal Neuron Importance
This work addresses a computational bottleneck for researchers and practitioners interpreting neural networks, offering a more efficient method, though it is incremental as it builds on existing attribution techniques.
The paper tackled the computational inefficiency of Total Conductance for measuring neuron importance in deep learning models by showing its equivalence to Path Integrated Gradients and providing a scalable implementation called Neuron Integrated Gradients, which was compared to DeepLIFT, finding DeepLIFT faster but lacking theoretical guarantees.
The challenge of assigning importance to individual neurons in a network is of interest when interpreting deep learning models. In recent work, Dhamdhere et al. proposed Total Conductance, a "natural refinement of Integrated Gradients" for attributing importance to internal neurons. Unfortunately, the authors found that calculating conductance in tensorflow required the addition of several custom gradient operators and did not scale well. In this work, we show that the formula for Total Conductance is mathematically equivalent to Path Integrated Gradients computed on a hidden layer in the network. We provide a scalable implementation of Total Conductance using standard tensorflow gradient operators that we call Neuron Integrated Gradients. We compare Neuron Integrated Gradients to DeepLIFT, a pre-existing computationally efficient approach that is applicable to calculating internal neuron importance. We find that DeepLIFT produces strong empirical results and is faster to compute, but because it lacks the theoretical properties of Neuron Integrated Gradients, it may not always be preferred in practice. Colab notebook reproducing results: http://bit.ly/neuronintegratedgradients