Skip to content

mrkcass/mechanizedAI

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

77 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

mechanizedAI

Mechanized AI, as it's name suggests is the application of AI by mechanical means. mechanizedAI is a framework of design and code on which useful AI can be built and deployed.

 

Somax

More concretely mechanized AI is Somax.

Somax adds sensors and motion to an AI inference and data collection engine. The sensors feed the AI with raw information. The AI uses the information to infer whether it should collect data or provide assistance. If data should be collected then sensors are directed by motion or other means and data is then captured. If instead assistance is needed motion or other means are used to provide the assistance directly or indirectly.

Hardware

Motion is provided by way of a 4 axis gimbal attached to a 3D color depth of field camera that can capture color images with per pixel depth at 1280x720 resolution at 30 frames a second. The gimbal has 800 steps per full revolution and completes a 90 degree directional change in the same time as a human. In addition, a 4 channel MEMs microphone array captures 48 khrz of synchronzied sample per microphone. Somax provides AI inference at a sustained maximum rate 400 GFLOPS Somax. Currently Somax operates off a single 12 volt 1.5 amp power supply. Future version will shed the on board ac supply in favor of a lower overall foot print using 4x 18650 cells and should be able to run all systems at full power for about an hour.

Software

Deciding what to do is an AI task called inference. Before inference can begin an AI must first be trained. Of the two, training is a much a bigger challenge than inference.

Training involves repeatedly passing millions of multidimensional floating point arrays as examples, for or against, though a compuation graph that usually has on the order of millions of nodes and weighted edges. After each example of data is passed through the graph, the weights of all edges going into all nodes are then, increased or decreased by a very, very small amount. this takes a lot computational power and lot of time. Training is usually the domain of GPU's which seem purpose built for the task. They speed things up but in the process use a lot of power. Most training these happens on NVidia 900 class or better hardware. At full power the cards easily consume 200 watts of power. the average phone has less than 4000 milliamp hours of power at under 5 volts of output. In other words you phone has enough power to run a a 900 class GPU, full out for about .2 seconds if it was even able to release all it power that fast in ther first place. In extreme cases this power draw must continue for days and even weeks.

Inference is another story and can easily be executed on a portable device in a reasonable amount of time. Somax uses 4 Myraid 2 compute engines like the one's found in the neural compute sticks. Each deleivers 100 GFLOPS of computation for 1 watt of power used. To give some scale, one Myriad 2 connected to a raspberry 3 increases it's computational power by 9x (movidius.github.io/blog/ncs-rpi3-mobilenets). So when the wind is blowing at our back, Somax should be able to process 36x more inferences in the same amount of time when compared to a raspberry pi 3. And for a bit more scale with respect spefically to inference, Somax will have that same ability as a Skylake Intel I7-6700k 4.2 GHz quad core desktop processor.

Mechanized AI will offer an end to end solution to address training and inference. The AI mech will infer, capture, and interact. When ready to train, the mech will attach to a DreamBox and begin training. the AI mech is always in charge of the training. The dreambox has no method to alter the mach AI and serves as the muscle for training. When training is complete, the mech will upload the new models, clean up, wake up and begin again with capture, infer, and interact functions. I hope to make this like the apple of AI devices. It does only a few things but it does them very well, without much help and everytime :-)

 

Resources

To learn more about mechanizedAI and Somax please visit: www.mechanizedAI.com

 

Licensing

Somax is my attempt to offer personal, private, and soley owned AI to everyone.

To that end there is no license for any design or implementation embodied within the domain mechanizedAI.

But there is a request.

Use it to make the world better for everyone :-)

 

-mechanizedAI