FiLM: Visual Reasoning with a General Conditioning Layer

FiLM introduces a general-purpose conditioning method for neural networks called Feature-wise Linear Modulation. A FiLM layer influences computation by applying a simple, feature-wise affine transformation to intermediate features, where the scaling and shifting parameters are produced from external conditioning information such as a question. This gives the network an explicit, lightweight mechanism to adapt its processing to the task at hand.

On visual reasoning, a task that requires multi-step high-level processing and has proven hard for standard deep learning, FiLM layers halve the state-of-the-art error on the CLEVR benchmark. The authors further show that FiLM modulates features in a coherent manner, is robust to ablations and architectural modifications, and generalizes well to challenging new data from few examples or even zero-shot, demonstrating the method's effectiveness and flexibility.