1. Implanted Devices for Human Machine Interface

- Minimum Invasive Implanted Chips

- Power Delivery and Communication based on

Ultrasound

- Multiple Channel Brain Dusts for Human Machine

Interface

3. Computer Architecture

- Any Novel Computing Mechanism that can Accelerate

>10 times

- High Performance AI Architecture for Self-driving

- Energy Efficient AI Architecture

- Sparse-aware AI Architecture

- Chiplet Solutions

2. Bio-plausible Novel AI Algorithms

- Bio-plausible Algorithms: Explainable and Robust AI

- Computer Vision for Self-driving: Classification,

Detection, Segmentation, PointCloud etc.

- Novel Sensor Fusion Models

- Compact Model Design: Pruning, Quantization,

Distillation etc.

4. AI Chips for Self-driving

- Video Processing Accelerator AI Chips for Real-time

Processing

- PointCloud AI Chips

- Secure Anti-adversarial AI Chip

- Online Learning AI Chip

Long-term goal: world smallest implanted dust

Function: to explore how the brain is working, especially for AI

Key technologies:

- Electric and magnetic sensors for sensing neural spiking activities

- 0 power communication technology

- 0 power data storage technology

- High density energy harvesting

- AI base reconstruction of the brain connections and activations

Competitors: Elon Musk's NeuraLink Company, which provides a Utah array like wird solution

Human Machine Interface – Brain Dusts:

- Extremely tiny – single neural size

- Wear a helmet outside the head as machine

- Multiple channel

- Can read >1 Million Neurons

- Can write to >10k Neurons

- Read / Write close loop <50ms

- Can read memory / write back memory

Tesla's accident was because the auto-pilot system regarded the white container as white cloud, thus the auto-pilot wants to go across under the truck. Why our brain does not make similar mistakes?

Once our brain learns an African Elephant, it can recognize the South Asian Elephant although the brain has never seen South Asian Elephant.

We want to borrow ideas from how the brain is working, to design better AI algorithms, especially for self-driving.

Long-term goal: brain-like algorithms

Key research focuses:

- Brain-like self-driving related algorithms

- Robust AI algorithms

- Explainable AI algorithms

- Common Sense

- Multiple source decision

- Energy efficient AI algorithms

The self-driving still needs time to mature, at 5-10 years, so we regard the next 5 to 10 years as golden age for self-driving research.

Long-term goal: L5 self-driving technologies, from algorithms to chipset

Key research focuses:

- Perception algorithms

- Robust AI algorithms

- Chipset design and fabrication

In our group, we main focusing on perception, which is the most computation expensive part.

A vertical solution is applied in the research, including robust models, compact model design, and AI architecture and chipset design and fabrication.

For technology roadmap for chips, we first focus on individual technologies – design tiny cores to process individual tasks using mature technologies like 55nm, and providing commercial IPs.

At the same time, the IPs are implemented in FPGAs, and deploy on real self-driving systems.

Once modules for single tasks are mature, we do fab SOCs to integrate things together, for commercial solutions.

Last but not least, we do Chipset / Cluster level solution for self-driving systems, compitable with IEEE car standards.