DARPA plans to launch a program aimed at developing biological chips that can handle low-power AI training and inference at the Edge.

The upcoming Organoid Cytomorphic Intelligence Resulting from Convergent Understanding and Information Transfer (O-CIRCUIT) program aims to develop unconventional biological processing units (BPUs).

In a special notice providing advance warning of the program, which has not been previously reported, DARPA states: "Modern computation involving AI training and inference has significant energy requirements, especially at the Edge. This affects everything from operational longevity to compute capacity and data transfer latency, which impacts warfighter performance at the Edge.

"O-CIRCUIT envisions unlocking operational advantages by developing unconventional biological processing units (BPUs) for Edge learning and inference with minimal power draw (mWh/day). O-CIRCUIT also envisions extending these BPUs to sensing physical stimuli and outputting instructions, such as directing drone navigation."

Such systems do not currently exist, although DARPA notes that nature has already proved it is theoretically achievable, with fly brains consisting of around 140,000 neurons while consuming only six mWh/day.

The program aims to demonstrate that it is possible to construct BPUs that have comparable size, weight, and power characteristics to naturally occurring neural structures to provide calibrated synthetic intelligence (training and compute memory) through convergent architecture and information management (connectivity and plasticity).

Over 42 months, O-CIRCUIT aims to develop a platform "at the confluence of organoid and synthetic biological intelligence where cell compositions, such as neural, glial, and immune cells, can be structured and composed (organically or synthetically) to perform complex training and inference tasks."

A demonstration challenge problem will involve developing trainable BPUs that "provide sophisticated inference and drone guidance via neural olfactory sensing to showcase an all-biological, converged synthetic intelligence sense-compute-act system."

Another task will be using the BPU to play a video game, with DARPA using Ms. Pac-Man as a potential example, at close to human-level proficiency.

A number of companies and research labs are developing biological processing units, aiming to merge the wet world of biology with that of semiconductors and servers.

FinalSpark develops living-neuron organoids attached to electrodes, which it uses for basic computing tasks. Developed from stem cells derived from human skin cells, the 'Neuroplatform' biocomputer is available on a research cloud platform.

Similarly, Cortical Labs operates its own cloud, with neurons 'living' on silicon hardware. In 2021, the company announced that the neurons were able to play Pong; this February, they linked 200,000 neurons together to play Doom.

Researchers at Johns Hopkins and Pomegranate Intelligence mapped a Drosophila (fruit fly) larva brain to a BPU, and were able to play chess.

However, such living systems are fragile and require controlled environments. Edge systems could face significantly more challenging conditions, requiring new research endeavors.