Ultra-low energy and area-efficient electronic systems are required to enable untethered computing at the edge of IoT. To realize self-learning edge-AI systems, conventional solely software-driven deep-learning neural networks becomes a major roadblock due the excessive energy expense of training. Hence, fundamental hardware change is likely needed. In this talk, we review our recent material innovations (E.g. Ferroelectric oxides and 2D Material) and we show how close coupling with new micro-architecture innovations (E.g. New memory physical layout and Monolithic 3D IC [1],[2](Fig.1(a)) may significantly accelerate in-memory computation. We explore wafer-level solution-processed CMOS-compatible use of 2D Material (MoS₂/WSe₂) [3](Fig.1(b)) to enable high-endurance memristors that can have properties superior to conventional oxide RRAMs.  We discuss the use and enabling of multi-gated HZO-based low-thermal-budget ferroelectric oxide memory transistors for new reconfigurable non-volatile logic and interconnect [4] (Fig. 2(a)). In co-operation with specific system-level innovations, we review material-system co-design in data encoding for deep convolution neural network. We show through material-device-aware data encoding, error correction, and novel physical memory layout (staggered + Manhattan arrays) [1], that aim to simplify in-memory data processing, one can significantly manage variabilities while accelerating convolution deep neural network operations and offer substantial low-energy opportunities towards reconfigurable Edge-AI systems. Extending similar innovations to photonics, we show that new ferroelectric-lithium niobate integration can enable reconfigurable photonics and in-memory compute for photonics [5] (Fig. 2(b)).