Our passion is to develop and deploy transformative perception and cognition tools, helping to create truly intelligent machines that serve society.

Software Perspective

ADC takes machine perception and cognition to a new level with a biomimetic design based on the mammalian extended visual pathway (EViP).  This unique approach includes:

*  Effective Unsupervised Learning (starting with a single sample per subject) to enable real-time, autonomous, adaptive capability to equip the dynamic perception. This is the front-end (“retina”) of the EViP architecture. It is viewed as short-term memory-like operation.

*  Self-Dynamic Supervised Learning based on self-evolving architecture to enable fast, adaptative, autonomous, constructively additive (no restart from scratch), and robust convergence learning, with much less computing power. This is the back-end (“visual cortex”) of the EViP architecture. This is considered long-term memory operation.

           Training Data             Classifiers

Intelligent learning system

It is a closed loop between short-term and long-term memory to enable intelligent operation. ADC is now equipped with a complete brain-like visual system, which comprises self-sample training data block and self-dynamic supervised learning block, to enable intelligent perception and cognition.

Hardware Perspective

Sparse Sensor (DARPA AIE IP2 Phase I and II)

EViP+ Gray-Scale Photo Array= Bio-Inspired Sparse Imager(BISI) (60x pixel intensity reduction)

Based on our studies for object detection, classification and tracking using BDD100K from UCB via DARPA AIE IP2 Phase I and Phase II program, the sparse image (61.5x pixel intensity reduction) (videos) is an equivalent performance with the full-color image (videos) based on the deep learning-based architecture (YOLO5) and it suggests the BISI carries sufficient salient features of each object interests.  This also reconfirmed the effective extracted features are used to compete and more advanced against our human visual systems from EViP.  We may wonder why EViP performs better human visual system, and the answer is that EViP is less sufficient in object salient extracted features as human visual system, but is beefed-up with 32- to 64- floating point accuracy against less than 8-bit accuracy from human, to over-perform in the bigger distractor size.

Massive Parallel In-Memory Learning and Processing Architecture

Regardless of parallel digital systems for learning, it is still sequential learning based on Von Neumann’s architecture and it is slow.

Massive parallel learning based on in-memory learning/processing architecture enables to inspire a brain-like learning architecture; hence it can very fast.  

In-Memory Processing Architecture

This work is done when Dr. Duong was with JPL/NASA.  The processing variations are the key challenges and re-corrected by a novel approach at ADC.

 

There are three directions in hardware, ADC is heading to:

• Sparse Sensor Implementation
• Massive Parallel Learning and Processing Device
• General Purpose High Speed In-Memory Processing Correlator (Compiler is expected to develop along with this)

The combination of software, hardware and its architecture enables us to position ahead in AI “True Intelligence” arenas.

 

 

MISSION

ADC’s mission is to become a leader in machines intelligence, perception, and cognition, spanning two domains: software and hardware.
In software, we are developing a self-training data generator and self-learning in the loop, to access the dynamic perception and cognition, enable an intelligence systems.
In hardware, we are developing a compact, low power, and high speed on-chip learning neural processor to serve as general purpose machine.