福岡工業大学大学院 / 工学研究科知能情報システム専攻
Research About In-memory Machine Learning Classifier
Scaling of memory technology increases the crisis of operating power and hardware variability in fields like Internet of Things and sensor networks, where the constraints of energy cost and hardware reliability are most rigorous. To get over such challenges, directions of compute memory and in-memory computation, where the computation is performed within memory (SRAM bit-cell), are becoming highlight recently. However such new computation structures can be vulnerable to acquired time-dependent variability for uncertain strength of variation. This work proposes the multiple-classifiers which correct not only circuit non-linearity but also timedependent variability within a range via employing several error-adaptive models pre-trained with different scaled variations.