<ul><li>Generative Binary Memory (GBM) is a novel pseudo-replay approach for Class-Incremental Learning (CIL).</li><li>GBM generates synthetic binary pseudo-exemplars using Bernoulli Mixture Models (BMMs).</li><li>The approach is applicable to any conventional Deep Neural Network (DNN) and supports Binary Neural Networks (BNNs) for embedded systems.</li><li>Experimental results show that GBM achieves higher accuracy than state-of-the-art methods on CIFAR100 and TinyImageNet datasets, and outperforms other CIL methods for BNNs with reduced memory usage.</li></ul>

Generative Binary Memory: Pseudo-Replay Class-Incremental Learning on Binarized Embeddings

Discover more