Artificial Intelligence

Somax Decision Streams

Decision Streams are sequences of operations that begin with data and end with action. It is easiest to imagine a Decision Stream as a physical flowing stream of water.

Like real stream, Decision Streams:

Have a source
Always flow in one direction.
Sometimes fork
Sometimes join
Sometimes pool
Eventually every drop of water ends up in the same ocean

Decision Streams follow the rules above and are built with multiple stages of the following primitive operations.

Data Capture. Somax has a wide array of data capture systems including cameras, microphones and environmental sensors. Data can also be captured from electronic community and public resources.
Pre-Classification Shaper. Shapers modify captured data so that it can be used by a specific Capture Classifier. For example some image classifiers may work better if it is given a black and white image instead of one in color. A shaper in this case would shape color images into black and white images.
Capture Classification. Somax employee's AI to interpret captured data in real time. AI's are loadable on demand and are tailored to a specific classification (or regression) task. AI's can be obtained from factory, community or commercial sources. AI's can also be constructed from scratch using a custom or pre-defined model with public and/or private data.
Post-Classification Translator. Translators take the output of a classifier and translates it to a form recognizable by an Application. A simple example would be taking a true or false output and converting it to a power on or power off message for an Application which remote controls a kitchen oven.
Classification Application. To be useful AI, Somax must be able to apply a classification to the world. Somax includes applications to apply these changes. Applications are wide and varied and can be acquired from multiple sources private and public. Somax factory install includes a variety of applications such as Camera-Lock_And_Hold and Camera-Scan_And_Hold which causes the camera to track an object in 3D space when it is triggered by a classification. Of course Yin must always have a Yang and so there is also a Do-Nothing Application which in some cases is actually the best action!

AI Intelligence Bubble....

"Hello again to all my friends I'm glad you came to play. Our fun and learning never ends. Here's what we did today."barney the dinosaur

If you understand at some level the methods above. If you at least kinda understand how they cause an action to occur when a condition is present, then congratulations!! You now understand neural networks !! Below is how some of the above concepts map to neural network concepts. Read on if you dare!!

Decision Stream = Computation Graphs. Here is Andrew Ng explaining Compution Graphs : https://www.coursera.org/lecture/neural-networks-deep-learning/computation-graph-4WdOY. The rules that govern how networks operate, those we liken to a stream of water, are expressed formally as the: architecture of Restricted Boltzman Machines, the Back Propogation algorithm and more specifically the Stochastic Gradient Descent algorithm.
Data Capture = Data sets. Wikipedia for Data Sets. Web page listing data set repositories.
Pre-Classification Shaper = Feature Engineering. Wikipedia for Feature Engineering. Stanford Crowd Scholar YouTube presentation on feature engineering.
Capture Classification = Inference. Nvidia article on the training and inference. Video of a neural network performing at type of inference called "semantic segmentation".
Post-Classification Translation / Classification Application = mechanizedAI. Neural networks stop short of actually doing something. However, it doesn't have to be the end of the story, at least not for those with mechanizedAI!

Adding and Creating Decision Streams

Decision Streams describe a process that begins with some type of data (images, videos, audio) as input and over any number or combination of operations, arrives at an action (change the camera orientation, display an alert.) The mechanizedAI framework will allow users to create, share and crowd source new Decision Streams.

Adding Decision Streams involves locating the stream, locating stream resources, building HotPlugAI's if needed and finally uploading it as a Stream Package to the Somax unit.

New Decision Streams can be original or improvements to existing streams or a combination of both. The extent to which you will need examples depends on what exists in the community and public resources versus how much variance can be found within the examples. This is sometimes referred to as the trade off between bias and variance.

Crowd Sourced Decision Streams are a way to kick start a new Decision Stream with limited examples. To crowd source, an inventor collects what examples are available and creates the intial stream configuration then submit's the low accuracy stream to a Somax crowd source engine. Members in the engine will allocate a certain amount of their Somax's resources to capturing data for crowd projects. The Somax Crowd Capture HotPlugAI will update routinely with new Classifications which will trigger the Crowd Capture Application to log and label a sample. Samples will be taken from normally captured data, and at the authorization of it's person, A Somax can can initiate crowd capture when capture services are idle. A user may option to allocate resources to specific crowd projects but must allow a small percentage for unspecified projects as well.

Imprinting

Imprinting is the emulation of mannerisms and responses consistent and acceptable to the imprinted user.

General Imprinting. General Imprinting is an imprinting that is the blend of an imprinted population. Somax will be factory loaded with a Human General Imprinting.
Specific Imprinting. Somax is designed to imprint on a user by learning and expressing similar traits such as body language or behavior patterns. For example if a user watches a scary movie with Somax. Somax will learn to react to scary movies in much the same way as the user.

Somax Hotplug AI

Somax provides artificially intelligent inference through factory, community, commercial and personally designed and trained Neural Network models.

Onboard Cached. A Somax unit will have many Hotplug AI's stored onboard for on demand load and classification.
Externally Fetched. With user authorization, Somax will also be able to fetch and load Hotplug AI's from external sources such as a personal device or the internet.
Community. Community HotplugAI's are the product of another Somax and human or groups of the same. Community HotPlugAI's will be available from public non-commercial sources. The HotPlugAI will be complete and correct solution but may require training or additional resources be gathered before the HotPlugAI will be usable. A HotPlugAI no matter what need be done or how long it may take will always have a single command, to make it so usable.
Commercial. Commercial HotPlugAI's are created and maintained by for profit and indirect for profit entities. The Community and Commercial labeling of a HotPlugAI is declared by the author and verified by the community. Commercial HotPlugAI's will always be accompanied by a Community Censor HotPlugAI to alert the community and user on and filter out personal data passed to and through a Commercial HotPlugAI. This will leverage a popular concept in commercial software called analytics which is used to measure software operating points and satisfaction. A community censor AI will leverage the same analytics to detect potential miss-use of personal intended and unintended data. Commercial HotPlugAI's can optionally obfuscate the output and implementation of a HotPlugAI and any subsequent Applications for the purposes of protecting intellectual property and or trade secrets.

"The brain has about ten thousand parameters for every second of experience. We do not really have much experience about how systems like that work or how to make them be so good at finding structure in data."

Geoffrey Hinton

"There is no sound ethical objection that can be made to honest, fair, and transparent commercial AI. There is also no logical reason for why it must be any other way."

mechanizedAI

Somax Community AI

Somax AI, at the discretion of an imprinted human can use and/or exchange data in an anonymous form with other Somax mechizedAI's for the purpose of improving Decision Streams.

The community is also responsible for a certain level of policing in support of personal protection.

Some services from a community will be in the form of data or applications. Most often a community will have one or more community oriented HotPlugAI's that provide and mange community responsibilities and services.

Thoughts and Ideas

Adversarial Inference and Motion Controllers

This is neat.

When thinking about a powerful compute engine, one must also think of power consumption. I was thinking about this with respect to the Myriad 2 vs the Myriad X. To get enough compute capacity I'll need more more than one Myriad each of which needs about a watt of power. Then I thought about the motors, each motor uses about 1 watt of power.

Then I realized a possible adversarial system seeking equilibrium. The equilibrium between motors that just want to run and compute engines that just want to think.

more to come, but this is a very interesting idea!

Learning Power Efficiency

Determining the voltage and current used at different points within the motor controller and digital electronics can be monitored at a sufficient sample rate (say twice the Nyquist rate for human head motion frequency) cheaply and simply.

This would be an excellent source of data for an AI to train on.

If we gave the AI the ability turn all the knobs on all the hardware on the Somax platform and said "minimize the power and current to all systems while maintaining a baseline performance" what might the network learn to do?

Could this be extended to add context, so that the settings used are based on power savings when the operational baseline is dynamic. For example, at night we can turn off or limit this use of the RGB camera (thus having a different operational baseline from daytime operation) without diminishing the quality of the data captured. Could we train a network to do this automatically?