FiLM: Frequency Improved Legendre Memory Model
Also known as: Frequency Improved Legendre Memory, FiLM Forecaster, Legendre Frequency Model, Frekans Tabanlı Legendre Bellek Modeli
FiLM is a long-term time-series forecasting architecture introduced by Tian Zhou and colleagues at NeurIPS 2022. It combines Legendre polynomial projections of the historical input with learnable frequency-domain filters applied to the resulting coefficient sequences. By representing history as a compact set of polynomial coefficients and filtering those coefficients in the frequency domain, FiLM enables efficient extrapolation over long prediction horizons without the quadratic cost of full self-attention.
Key highlights
- Linear computational complexity in sequence length, enabling very long look-back windows
- Compact Legendre memory compresses history without information loss up to order K
- Frequency-domain filtering provides an interpretable mechanism for denoising
- Competitive long-horizon forecasting accuracy with a simple, lightweight architecture
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
FiLM suits long-horizon univariate and multivariate time-series forecasting tasks where the series exhibit trend and periodic structure. It is a strong baseline when Transformer-based models are too expensive or when input sequences are very long. It assumes stationarity or slow non-stationarity and works best when dominant frequencies are stable across the look-back window. For highly irregular, event-driven, or non-stationary series, recurrent or diffusion-based approaches may be preferable.
Strengths & limitations
- Linear computational complexity in sequence length, enabling very long look-back windows
- Compact Legendre memory compresses history without information loss up to order K
- Frequency-domain filtering provides an interpretable mechanism for denoising
- Competitive long-horizon forecasting accuracy with a simple, lightweight architecture
- Polynomial order K is a sensitive hyperparameter; too low loses detail, too high overfits
- Assumes that dominant frequencies in the look-back window are representative of future patterns
- Limited capacity to model abrupt distributional shifts or sudden regime changes
- Multivariate channel mixing is basic compared to dedicated cross-variable attention mechanisms
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How does FiLM differ from FEDformer and Autoformer?
FEDformer and Autoformer retain a Transformer backbone with frequency-enhanced or autocorrelation-based attention. FiLM replaces the attention mechanism entirely with Legendre polynomial projection and frequency filtering, yielding a simpler architecture with linear complexity and no attention computation.
What is the role of Legendre polynomials in FiLM?
Legendre polynomials form an orthogonal basis on a bounded interval. Projecting the look-back window onto this basis compresses the input into K coefficients that capture different temporal scales. This projection is lossless up to the chosen order and provides a natural analytical form for extrapolation beyond the observed window.
Does FiLM require GPU acceleration for practical use?
FiLM's linear complexity means it can run on CPUs for moderate sequence lengths and forecast horizons. However, GPU acceleration is recommended for training on large datasets or when using high polynomial orders K, as the frequency-domain matrix operations benefit significantly from parallel hardware.
Sources
- 1.Zhou, T., Ma, Z., Wen, Q., Sun, L., Yao, T., Yin, W., & Jin, R. (2022). FiLM: Frequency improved Legendre memory model for long-term time series forecasting. NeurIPS.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 2). FiLM. ScholarGate. https://scholargate.app/deep-learning/film