Shallow-Deep Networks: Understanding and Mitigating Network Overthinking

TL;DR

This paper investigates the problem of overthinking in deep neural networks, introduces the Shallow-Deep Network (SDN) for internal prediction transparency, and demonstrates how confidence-based early exits can significantly reduce inference costs while maintaining accuracy.

Contribution

The paper proposes SDNs with internal classifiers for better prediction transparency and introduces confidence-based early exits to mitigate overthinking, reducing inference costs by over 50%.

Findings

SDNs can effectively mitigate overthinking in DNNs.

Early exits reduce inference cost by more than 50%.

Overthinking can cause correct predictions to become misclassifications.

Abstract

We characterize a prevalent weakness of deep neural networks (DNNs)---overthinking---which occurs when a DNN can reach correct predictions before its final layer. Overthinking is computationally wasteful, and it can also be destructive when, by the final layer, a correct prediction changes into a misclassification. Understanding overthinking requires studying how each prediction evolves during a DNN's forward pass, which conventionally is opaque. For prediction transparency, we propose the Shallow-Deep Network (SDN), a generic modification to off-the-shelf DNNs that introduces internal classifiers. We apply SDN to four modern architectures, trained on three image classification tasks, to characterize the overthinking problem. We show that SDNs can mitigate the wasteful effect of overthinking with confidence-based early exits, which reduce the average inference cost by more than 50% and preserve the accuracy. We also find that the destructive effect occurs for 50% of misclassifications on natural inputs and that it can be induced, adversarially, with a recent backdooring attack. To mitigate this effect, we propose a new confusion metric to quantify the internal disagreements that will likely lead to misclassifications.