New advanced processors from IBM mimic the mammalian brain -- let's hope they're on our side

International Business Machines, Inc. (IBM) has been working on a project co-funded by both the U.S. military and various academic partners to develop a chip that "thinks" and "processes" perception like a fleshy life form.  IBM calls the project SyNAPSE (Systems of Neuromorphic Adaptive Plastic Scalable Electronics) after the structure in neurons that's used to signal the neurons they're associated with.

I. Meet the First Full-Fledge Neural Networks Dev Kit

IBM research fellow Dr. Dharmendra S. Modha has partnered with a colleague in academia -- Professor Rajit Manohar a VLSI expert at Cornell University -- and iniLabs, Ltd, a spinoff of the Universität Zürich's (Univ. of Zürich)'s Institute of Neuroinformatics to develop the "thinking" chip.

Fifty-three million dollars in grant money later, they have produced a complete neural networks development packages that aims to offer some abstraction and to ease the process of developing neuronal networks that can process and respond to sensor input.

The basic offering consists of:
  • Hardware Unit
  • SDK
    • Used to programmatic network of the fundamental units
    • Maps unit interactions
    • Maps I/O to network
  • Examples/Library
    • 150+ premade corelet examples
    • Real-time actuator and sensor examples are included
  • Simulator
    • Sort of like smartphone development "simulated hardware"
    • Uses software model of hardware unit to predict what your neural network code will do
    • Allows you to get network basically working before loading it onto actual hardware
    • Saves the cost of having to have the development hardware up front
  • Laboratory
    • Bundled package containing SDK, simulator, hardware support, and examples library
One example showcased by IBM shows a neural network using a retinal sensor, which mimics the human eye.  A real human brain translates over 1 terabyte of raw visual data a day into recognized shapes, tracked motions, depth perceptions, and self-tuning feedback to the sensors based on light conditions.  By comparison IBM's network is far simpler, accomplishing basic shape detection.

Retinal sensor
(left: retinal sensor "vision"; right: significant shape output)

Dr. Modha comments, "Architectures and programs are closely intertwined and a new architecture necessitates a new programming paradigm.  We are working to create a FORTRAN for synaptic computing chips. While complementing today’s computers, this will bring forth a fundamentally new technological capability in terms of programming and applying emerging learning systems."

II. Project Motivations -- Killer Drones, Self-Driving Cars, and Stock Picking

The project is funded by Defense Advanced Research Projects Agency (DARPA).  DARPA typically does not directly participate in the non-military projects it funds, but more often than not it funds projects that it considers of interest to the future technological progress of the U.S. military and intelligence community.

Neural networks in the future could be used to create fully autonomous attack drones and ground-based war robots.  Facial recognition neural networks could also sort through feeds of small ground based and high-flying high-resolution air based camera drones, as well as hacked cloud-connected cameras in the target state, to recognize targets slated for termination.  With the target's location in hand, the killing robots could then be dispatched using lightweight neural networks to perform target identification and aiming.

With both domestic armed drone use and federal camera surveillance of citizens on the rise, similar tactics could be applied domestically against individuals the ruling administration deems "terrorists."

Predator missile
Drones could use multi-sensor networks to locate targets, then employ lightweight neural networks to locate their target and aim the killing shot. [Image Source: Drone Wars UK]

Currentlty the project is in its third phase, which received $12M USD in additional funding.  IBM hosts details of the past rounds (Phase 0Phase 1, and Phase 2), for those interested.

IBM and its academic partners are more interested in the science and financial implications of the project than the warfare side.  Among the applications they're eyeing are:
  • predictive stock/currency trading
  • drug development
  • climate/environmental monitoring
  • autonomous cars
  • service robots
  • behavioral marketing
A graphical representation of IBM's neural network abstraction scheme

IBM describes the justification for this new computing paradigm commenting:

Although they are fast and precise “number crunchers,” computers of traditional design become constrained by power and size while operating at reduced effectiveness when applied to real-time processing of the noisy, analog, voluminous, Big Data produced by the world around us. In contrast, the brain—which operates comparatively slowly and at low precision—excels at tasks such as recognizing, interpreting, and acting upon patterns, while consuming the same amount of power as a 20 watt light bulb and occupying the volume of a two-liter bottle.
IBM’s long-term goal is to build a chip system with ten billion neurons and hundred trillion synapses, while consuming merely one kilowatt of power and occupying less than two liters of volume.

Neural networks are the most optimal design ever produced by nature for the purpose of survival by dramatic transformation of the environment in which a living organism lives.  Neural networks could one day help machines serve mankind performing menial tasks and helping to cure disease.  Alternatively, they could potentially be the most lethal combination of destructive power, obedience, and precision ever witnessed by mankind, potentially killing millions with the press of a single button.

In that regard neural networks and ubiquitous massively parallel sensor networks are perhaps the most promising yet deadly tool produced by mankind since the fateful splitting of the atom.

Source: IBM

"And boy have we patented it!" -- Steve Jobs, Macworld 2007

Most Popular Articles

Copyright 2018 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki