ANP-O: A 67 <i>μ</i> W/Channel, 0.13 nW/Synapse/Bit Nose-on-a-Chip for Noninvasive Diagnosis of Diseases With On-Chip Incremental Learning

Portable electronic nose (E-nose) is considered as an innovative diagnostic tool designed to detect pathological changes in the body by analyzing a patient’s exhaled breath. However, the accuracy of E-nose is affected greatly by environmental factors and the use of different devices in various settings, such as hospitals, health centers, or homes. To address this issue, we propose a$67{~\mu }$W/Channel E-nose for noninvasive diagnosis of diseases with on-chip incremental learning. To our knowledge, this is the first E-nose system which supports on-chip learning. It has three features: 1) on-chip incremental learning enables retraining the network rapidly and locally to overcome accuracy degradation caused by environment with changing conditions and patients using different devices; 2) an ADC-free front-end achieves faster response and better power efficiency without the need for a conventional ADC conversion stage; and 3)event-driven asynchronous logic effectively avoids the power brought by a global clock. Besides, we propose a self-adaptive synapse weight update skipping (SWUS) mechanism and a trained synapses weight low-width storage (TWLS) method to eliminate 82% redundant synapse weight update, and 50% storage cost of the trained synapse weights with no accuracy loss simultaneously. Benefitted from these methods, the E-nose system obtains a 98% accuracy for 10-class gas recognition with one-shot learning and a 0.13 nW/Synapse/Bit power density, achieving over 6% and$8.0{\times }$improvement, respectively, over the state-of-the-art (SOTA) gas recognition chips.