Building Neural Network Models That Can Reason

Deep learning has had enormous success on perceptual tasks but still struggles in providing a model for inference. To address this gap, we have been developing networks that support memory, attention, composition, and reasoning. Our MACnet and NSM designs provide a strong prior for explicitly iterative reasoning, enabling them to learn explainable, structured reasoning, as well as achieve good generalization from a modest amount of data. The Neural State Machine (NSM) design also emphasizes the use of a more symbolic form of internal computation, represented as attention over symbols, which have distributed representations. Such designs impose structural priors on the operation of networks and encourage certain kinds of modularity and generalization. We demonstrate the models’ strength, robustness, and data efficiency on the CLEVR dataset for visual reasoning (Johnson et al. 2016), VQA-CP, which emphasizes disentanglement (Agrawal et al. 2018), and our own GQA (Hudson and Manning 2019). Joint work with Drew H