Fig. 3: HMN for MCL tasks.

a Architecture of HMN, consisting of an ANN-based backbone network to extract task-level information and an SNN-based branch network to perform specific tasks. Learnable HUs are adopted for adaptively adjusting modulation signals to meet the requirements of the branch network. The HUs are installed at the backend of the backbone network (Fig. 1e-(1)), and are jointly trained by the task similarity objective. b Similarity matrix of tasks, calculated based on the Hamming distance between the permutation indices of tasks in this experiment. c, d The correlation matrix of mean activations of hidden neurons of the HMN and the single SNN, respectively. e T-SNE embeddings of the sample-specific modulation signals generated by the backbone network. The triangle denotes the modulation signal of the test samples of the backbone network. f Mean accuracy of different models on unlearned tasks with different levels of similarity. g Mean accuracy of the ANN-only model and the SNN-only model with a similar architecture of HMN, respectively. h The first 128 dimensions of the task-specific modulation signals before thresholding. i Mean accuracy of different models on learned tasks versus the number of learned tasks. All the experiments in this demo are independently run four times with the same setting. The error bar represents the standard deviation of the results.