Associative Recurrent Bilinear Optimization for Domain-Generalized Binary Neural Networks

Binary Neural Networks (BNNs) show great promise for resource-constrained embedded devices. However, BNNs with degraded representation possess less domain generalization (DG) capability in real-world scenarios. In this work, we observe that conventional DG methods are ineffective in pursuing the flat minimum for BNNs, which is primarily caused by the sign function. Furthermore, existing BNNs neglect the intrinsic bilinear relationship of real-valued weights and scaling factors, resulting in an ineffective optimization process. To address this issue, an Associative Recurrent Bilinear Optimization method for BNNs (ARBONNs) is proposed to improve the performance for DG tasks. Specifically, we introduce a recurrent optimization scheme and Associative Density-ReLU to sequentially backtrack the sparse latent weights, which utilize flat minimum and the density of bilinear variables as conditions to decide whether the backtracking operation is performed or not. In this way, ARBONNs can achieve a generalized solution of a flat minimum with a controllable learning process. Moreover, we introduce a domain invariant module (DIM) to effectively mitigate the impact of domain diversity on activation binarization. Our ARBONNs show impressive performance over state-of-the-art BNNs on various models on various DG tasks. As a generic method, ARBONNs also achieve significant performance improvement on the in-domain tasks.