Wednesday, March 21, 2018

Forbes: Is Deep Learning The Big Bang Moment For AI? (English)

Форматирование не сохранено.

Very interesting. Article in Forbes about state-of-the-art in Machine Learning. Not in professional literature, Forbes!

It could be argued that Geoffrey Hinton’s recent success with neural networks is the Big Bang moment for artificial intelligence. Deep learning has enabled today’s AI systems to beat Go world champions and translate data into innovation for industries like finance and energy. But the field still struggles in areas requiring broader intelligence, and the question remains whether a series of incremental innovations to the current foundation will lead AI to a new level of sophistication -- one that can outpace human ingenuity.

Current AI Can Dominate In A Game Of Go But Is Flustered By Nursery Rhymes

Today, the AI research agenda is focused on deep learning, which processes large data sets to solve narrow and specific tasks at hand. Deep neural networks can learn complex functions to solve intricate problems but only within certain parameters. For example, Google's AI-based AlphaGo is easily able to beat the best human player at the incredibly sophisticated game Go because it is played using set rules.

But outside of situations with set parameters, the real question is if data-driven AI can outperform humans in the complex world we live in.

Deep learning is very good at solving a particular set of problems but does not directly address many important issues, such as long-term planning. While deep learning can solve the problem of finding human experts for time-consuming data input by automating perception and knowledge acquisition, there is more work required to recreate human-like semantic understanding for complex actions.

Natural language processing is an excellent example of this. Though the complexities of language seem beyond statistical correlation, machine learning researchers have successfully used their techniques to manipulate language to complete tasks. However, the techniques don't extend to natural language understanding -- to a computer, this is just another algorithm, not a sentence with meaning.

To become genuinely intelligent, AI systems need to have human-like reasoning skills combined with a machine scale to process data. Consider the nursery rhyme “Jack and Jill went up the hill to fetch a pail of water.” A modern system can answer, “Where did Jack and Jill go?” AI is learning to answer questions like, “Are Jack and Jill still at the top of the hill?” demonstrating inductive reasoning. However, an AI system would be unable to describe how Jack and Jill retrieved the water, as it requires background information not specified in the rhyme.

If This Is AI’s Rebirth, What Will AI Grow Up To Be?

Ниже есть продолжение.

The next frontiers for AI researchers are matching humans’ abilities in abductive reasoning and building a symbolic structure. As deep learning is integrated with other techniques available within AI, it will serve as a key foundation for another set of abilities, namely working with smaller data sets, automating learning processes and increasing unsupervised learning.

The first problem to tackle is the requirement of big data sets. With deep learning, large amounts of data ensure accurate answers, but it’s not feasible to have an expansive data set for many real-world use cases (for example, clinical trials). Techniques on the research agenda (like active learning, zero-shot learning and context framing) are showing promise in this arena.

Unsupervised learning by AI systems is also a hot topic on research agendas. The route to fully autonomous systems will have multiple stages, each moving toward human-like learning. For example, this can be seen with demonstrative learning, where a robot can learn from “watching” a task performed by humans or videos.

As AI moves toward unsupervised learning and broader uses, it must become explainable. The models that neural networks produce don’t necessarily make logical sense (hence, how AI can make moves in Go that a human wouldn’t think of?). When using AI to, for example, plan maintenance for an oil rig, where going offline costs millions of dollars per day, companies expect a clear, verifiable explanation as to why AI is recommending maintenance. Progress in this field is advancing quickly, and as AI demonstrates competence, it will likely gain buy-in for everyday uses. Significant research is focused on experimenting with the architecture of neural networks to provide transparency.

Dr. Manuela Veloso, head of the machine learning department at Carnegie Mellon University, describes the future of the relationship between humans and AI as “symbiotic autonomy.” She envisions research on methods of correction to be incorporated in the AI machinery. Thus, instead of telling Amazon’s Alexa explicitly, “Stop playing music,” the system will learn the context of the situation, so by saying, “I’m leaving now,” it will know it should turn off the music.

Deep learning is promising to provide a mathematical representation of the universe, and its potential is still in its infancy. The economic growth from the rebirth of AI is spurring researchers to expand its abilities -- but in the direction of improvement rather than lofty goals of general intelligence.

Only time will tell whether deep learning’s recent success is the Big Bang moment for AI, but given that it is fueling the field’s growth and evolving at rapid speed, it certainly looks that way.

No comments:

Post a Comment