early exit networks

**Early Exit Networks** are **neural networks with intermediate classifiers at multiple layers that allow easy inputs to exit early** — if an intermediate classifier is confident enough, the remaining layers are skipped, saving computation for simple inputs while using the full network for difficult ones. **How Early Exit Works** - **Exit Branches**: Attach classifiers (small heads) at intermediate layers of the network. - **Confidence Threshold**: If an exit branch's confidence exceeds a threshold $ au$, output that prediction. - **Skip Remaining**: All subsequent layers and exits are skipped — computation savings proportional to exit position. - **Training**: Train exit branches jointly with the main network, balancing all exit losses. **Why It Matters** - **Adaptive Compute**: Easy inputs use less computation — average FLOPs per sample decreases significantly. - **Latency**: In real-time systems, early exits guarantee latency bounds — hard cases are truncated. - **Edge Deployment**: Enables deploying large models on edge by averaging less computation. **Early Exit Networks** are **fast-tracking the easy cases** — letting confident intermediate predictions bypass the remaining computation.