Prompt Description

Figure Title: M=2 Co-teaching Time Series Prediction Training Framework Overall Layout: The diagram is arranged from left to right, containing four main stages: Input → Dual Model Parallel Prediction → Small Loss Sample Selection → Cross Update. The middle part consists of two symmetrical model branches with identical structures but independent parameters. 1. Input Module (Leftmost): Draw an input block, labeled as: Time Series Input Window Description within the block: Construct samples using a sliding window. Input length is 𝐿 , prediction length is 𝐻 Comment below the block: Keep the input sequence clean. Supervisory noise is only injected into the prediction target. Draw two arrows from this input block, pointing to the upper and lower model branches, respectively. 2. Dual Model Parallel Structure (Middle): Draw two model blocks arranged vertically, identical in size: Upper Model: Model 𝑓 𝜃 1 Lower Model: Model 𝑓 𝜃 2 Comment within the model blocks: Same architecture Parameters are independent of each other Each model block receives the time series window from the input module. 3. Parallel Prediction and Loss Calculation: Draw arrows from each model block to the right, connecting to the corresponding prediction output block: Upper Prediction Block: Prediction Output 𝑌 ^ ( 1 ) Lower Prediction Block: Prediction Output 𝑌 ^ ( 2 ) Below each prediction output block, draw a loss calculation module: Module Name: Window-level Loss Calculation Description within the module: Aggregate errors over all time steps within the prediction window. Use a basic loss function (such as MSE or Huber). Obtain the window-level loss value for each sample. 4. Small Loss Sample Selection: Draw arrows from each "Window-level Loss Calculation" module to the right, connecting to the sample selection module: Upper Selection Block: Small Loss Sample Selection (Top r%) Lower Selection Block: Small Loss Sample Selection (Top r%) Description within the module: Sort samples according to window-level loss. Select the top r% of samples with smaller losses. Consider their supervisory information to be relatively reliable. 5. Cross Sample Exchange Mechanism (Key Part): Draw cross arrows between the two "Small Loss Sample Selection" modules: From the upper selection module pointing to the lower model. From the lower selection module pointing to the upper model. Label next to the cross arrows: Cross Update Text Description: Each model does not use its own selected samples for updating. Instead, it uses the samples selected by the other model to update its parameters. Avoid self-reinforcement of the model on noisy samples. 6. Parameter Update: Connect the cross arrows back to the corresponding model blocks: Model 𝑓 𝜃 1 : Uses the samples from model 𝑓 𝜃 2 for parameter updates.