Artificial Intelligence: What’s Wrong, What’s Missing, What’s Next?

This is the first part of a serial three-part blog on Artificial Intelligence. Part one is What's Wrong?, part two is What’s Missing?, part three is What’s Next??This is Part One:

Artificial Intelligence: What’s Wrong?

In spite of the hype, persistent problems with Artificial Intelligence are blocking progress.

In the past decade, VCs and companies have spent $100+ billion on neural networks and machine learning and have only managed to move the needle about 10%.?Progress toward the ultimate objective of human-level general intelligence or better has stagnated – basically no progress.?The current neural network back-propagation training methods have hit a wall.?Throwing oceans of data, warehouses of super-computers, and months of training time at the problem has NOT worked – pushing real hard is NOT working.?Intelligence at any level is more complex than anticipated.?Something different is needed: a fresh look into AI fundamentals, and a new technical approach are needed.

What’s Wrong?

Existing AI systems might as well be toasters.

Today’s neural nets are so brittle that no one truly believes that they understand or know anything, nor can anyone safely depend on them.?They are statistics engines, big data systems, software tools, nothing more.?They can’t and don’t function independent of people.?They execute statistical algorithms; they have no awareness, no understanding, no context, no meaning, and they can’t think.?They have no common sense, or understanding of causality.?They cannot generalize from one task to another.?They can’t learn from experience – they don’t experience.?Robots are not coming for anyone’s job anytime soon.

A neural net trained to recognize sheep, cannot recognize cows, or horses, or chickens, or you, or me, or your cat.?It would have to be re-trained to recognize your cat, and then it would forget the sheep.?AIs have a difficult time picking out your cat from all other cats.?AIs have no real idea what sheep are, or what cats are.?Researchers have demonstrated that trained neural nets see sheep in images that look like video noise – no discernable images of anything at all.?Neural nets do not see sheep the way people do, perhaps what they see in sheep images looks like video noise.?How does that make sense??How reliable is that??Is this sheep-seeing neural net to be trusted to see and count sheep??Would you let such a neural net drive your car??Ford Motors recently announced that they were delaying the release of their RoboCar for five or more years!?Why is that??Fatal crashes involving RoboCars on autopilot perhaps??

Ford Motors recently announced that they were delaying the release of their RoboCar for five or more years!?

Even worse, neural nets are black boxes.?Neural nets cannot explain why they make the decisions that they make; they cannot provide reasons or evidence for what they do.?At least a person can make up some possibly plausible story, or point to relevant data, about why they do what they do.?Neural nets can’t even do that.?Neural nets cannot provide meaning to affirm their results.?Neural nets are unaware when they make mistakes.?It’s people who take neural net results as meaningful, and it’s people who will suffer from their all too inevitable mistakes.?Automated systems that unknowingly make the same mistakes thousand times per second worry me.?Not to mention security.?Adversarial attacks on neural nets have easily misled them.?If AI’s cannot protect themselves, how can they be safe enough for people to trust?

Current neural net technology is complex and not reliable.?There’s a veritable zoo of network architectures to choose from when constructing a new neural network.?Setting up and configuring a neural network is complex and is not guaranteed to work as expected.?Weeks of computation training a new network are easily wasted when the neural net does not converge to working solution – a trained net that performs up to expectation.?Then it’s back to square one, choosing a different set of initial configuration parameters and perhaps a new architecture, then running the training over again – rinse and repeat.?It has been difficult for different teams using slightly different open source tools to replicate and corroborate each other’s results.?Retraining in production AI systems is essential to keep the system up-to-date with the latest real-world data, but a single false step in the training data can ruin the system’s performance.?Good performance and good results are not guaranteed!

The Good News: The promise of artificial intelligence is so enormous, no one is giving up.

The current state-of-the-art in AI is among the highest of human technical achievements.?People working on neural net technology have the most exciting jobs on the planet.?AI systems are amazingly interesting, exciting toys, able to best humans on many difficult recognition tasks and games.?Yes, training AIs requires oceans of data, warehouses of super-computers, and weeks of computer-time; but even with these issues AI is sweeping the industry helping to automate everything.?Big Data and Big Automation are Big Business, and getting bigger every day.?Every company in the world is climbing on the AI bandwagon – no one wants to miss out this time, like they missed out on the Internet and mobile revolutions.?But Big Automation is deeply limited, fragile, not aware that it has biases, and not aware when it makes errors.?Yet nevertheless, full steam ahead.?Neural net automation, as powerful as it is, requires significant human input and oversight.?Humans need to spot check and error correct Big Data results.?Big Data analytics is very powerful, growing, creating huge new industries, and it's lots better than what came before.

The worry about super intelligent AIs wiping-out humanity or robots coming for people’s jobs anytime soon is greatly exaggerated.

This is only the start.?Smarter AI could do lots, lots more.

This is the first part of a serial three-part blog on Artificial Intelligence. Part one is What's Wrong?, part two is What’s Missing?, part three is What’s Next??

About the Author:

Bruce Amberden is a visionary, inventor, designer, architect, startup founder, CTO, and Engineering VP.?He has founded startups and worked with leading technology companies in Silicon Valley to create amazing software products.?Bruce has over 20 years of experience as a software engineer and as a leader inspiring tiger teams to do brilliant innovative work.?He has a Masters of Science in Physics and is an Armchair Astronomer.?Bruce is working on a personal project building breakthrough semantic AI technology.?Bruce is seeking new employment and is available for startup and technology consulting.