Lightscline's manifesto

Lightweight large-scale AI for the physical world

We are very selective with our attention. We know that the brain can easily get overwhelmed if we focus on several things simultaneously. Moreover, the brain has a very small power and memory budget, and needs to manage cognitive, biological, and motor tasks, forcing it to be compute efficient in its predictions.

Now, consider the embedded devices/computers interacting with the physical world as a proxy for the brain:

1. In space: Satellites collecting & analyzing multi-spectral images

2. In air: Drones & UAVs collecting images for real-time object detection

3. On land: Sensors on wind turbines, steam engines, pumps, and motors

4. In wearables: Operate in <256 kB of SRAM and ~mW power consumption

5. In sea: Unmanned underwater vehicles mapping the sea floor and looking for hidden objects

In all these scenarios, the processor needs to operate in resource constrained environments similar to the brain. These often look like:

(i) mass limits for processors in orbits, every additional kg means ~$1000s of added costs

(ii) <256 kB SRAM, micro-milli second prediction latencies, ~mW power consumption

(iii) intermittent or unavailable connectivity for remote operations

(iv) 40+ hours of manual analysis for every hour of data collected

(v) 73+ trillion GB of sensor data by 2025

Shouldn’t then the computing algorithms used in these environments be as efficient if not more than the human brain? What insights can we leverage to obtain 100x improvements in the design of post Nyquist analysis algorithms?

Welcome to Lightscline.

By leveraging the foundational insight that real-world sensor data is very redundant, Lightscline AI reduces 90% of the AI infra and people related time and costs by selectively focusing on the 10% important data. By automatically identifying the 10% important data using a learning process, we enable 10-74x reduction in compute, power, storage, transmission, and latency requirements, enabling several applications currently unobtainable due to SWaP-C constrains. For eg: running AI inference on <256 kB SRAM and ~mW power footprint.

Our mission is to create lightweight large-scale intelligence, primarily at the intersection of the digital and physical worlds (embedded AI). We believe that the next great platform companies will be building AI software for hardware. This is how we will enable millions of wearables, lunar and Martian communications, millions of humanoid robots, swarms of drones and UAVs, smart home devices, non-invasive personal medical devices, spatial computing, etc., creating immense lifestyle improvements and GDP growth.

Lightscline’s selective AI will also revolutionize scientific computing by creating post-Nyquist techniques that enable dataset prioritization for tasks like material discovery and spectroscopy.

You can read more about us here.