Imagine learning how to ride a bicycle! You learn to balance - pedal - ride on a straight line - turn - ride in busy streets - All set!!! It takes step by step learning & then if you are offered a different bicycle you would try to apply the “truths” you discovered in your earlier learning process & quickly pick up the new one too. Our machines so far perform the tasks they are programmed for and as obedient followers carry out the required job. However, the new wave of technology is striving to make the machines more intelligent, to not only seek but offer assistance, to make our decision making better, help an ageing population store & retrieve memories that fade and much more!!! Sounds interesting? Conniving? …???
Vishal Dhupar, Managing Director – Asia South at Nvidia would be discussing Re-Emergence Of Artificial Intelligence Based On Deep Learning Algorithm as part of the invited keynote on Day 1 DVCon India 2017. Passionate about the subject, Vishal shares the background & what lies ahead for us in the domain of AI & Deep Learning. Extremely useful from beginners to practitioners!!!
Vishal your keynote focusses on AI & Deep learning – intricate & interesting topic. Tell us more about it?
Curiously, the lack of a precise, universally accepted definition of AI probably has helped the field to grow, blossom, and advance at an ever-accelerating pace. Claims about the promise and peril of artificial intelligence are abundant, and growing.
Several factors have fueled the AI revolution which will be the premise of my talk. Touching upon how machine learning is maturing, and further being propelled dramatically forward by deep learning, a form of adaptive artificial neural networks. This leap in the performance of information processing algorithms has been accompanied by significant progress in hardware and software systems technology. Characterizing AI depends on the credit one is willing to give synthesized software and hardware for functioning appropriately and with foresight. I will be touching upon a few examples of AI advancements.
How do we differentiate between machine learning, artificial intelligence & deep learning?
Machine learning, deep learning, and artificial intelligence all have relatively specific meanings, but are often broadly used to refer to any sort of modern, big-data related processing approach. You can think of deep learning, machine learning and artificial intelligence as a set of concentric circles nested within each other, beginning with the smallest and working out. Deep learning is a subset of machine learning, which is a subset of AI. When applied to a problem, each of these would take a slightly different approach and hence a delta in the outcome.
Artificial Intelligence is the broad umbrella term for attempting to make computers think the way humans think, be able to simulate the kinds of things that humans do and ultimately to solve problems in a better and faster way than we do. Then, machine learning refers to any type of computer program that can “learn” by itself without having to be explicitly programmed by a human. Deep learning is one of many approaches to machine learning. Other approaches include decision tree learning, inductive logic programming, clustering, reinforcement learning, and Bayesian networks. Deep learning was inspired by the structure and function of the brain, namely the interconnecting of many neurons.
Some of the discussions on deep learning are intriguing. Does it lead to machines taking over jobs?
Machines are getting smarter because we’re getting better at building them. And we’re getting better at it, in part, because we are smarter about the ways in which our own brains function.
Despite the massive potential of AI systems, they are still far from solving many kinds of tasks that people are good at, like tasks involving hand-eye coordination or manual dexterity; most skilled trades, crafts and artisan- ship remain well beyond the capabilities of AI systems. The same is true for tasks that are not well-defined, and that require creativity, innovation, inventiveness, compassion or empathy. However, repetitive tasks involving mental labor stand to be automated, much as repetitive tasks involving manual labor have been for generations.
Let me give you an example your calculator is smarter than you are in arithmetic already; your GPS is smarter than you are in spatial navigation; Google, Bing, are smarter than you are in long-term memory. And we're going to take, again, these kinds of different types of thinking and we'll put them into, like, a car. The reason why we want to put them in a car so the car drives, is because it's not driving like a human. It's not thinking like us. That's the whole feature of it. It's not being distracted, it's not worrying about whether it left the stove on, or whether it should have majored in finance. It's just driving.
What are the domains that you see would see faster adoption & benefits of these techniques?
In healthcare, deep learning is expected to extend its roots into medical imaging, translational bioinformatics, public health policy development using inputs from EHRs and beyond. There is rapid improvements in computational power, fast data storage and parallelization have contributed to the rapid uptake of deep learning in addition to its predictive power and ability to generate automatically optimized high-level features and semantic interpretation from the input data.
Seems like the ASIC design flow/process can be equally benefited from these techniques. Your views on it?
Deep Learning in its elements is an optimization problem. Its application in any work flow or design process where there is scope for optimization carries enormous benefits. With respect to the design, fab and bring up of ICs, deep learning helps with inspection of defects, determination of voltage and current parameters, and so. In fact, at NVIDIA we carry out rigorous scientific research in these areas. I believe as we unlock more methods of unsupervised learning, we’ll discover and explore many more possibilities of efficient design where we don’t entirely depend of large volumes of labelled data which hard to get in such a complex practice.
What are the error rates in execution we can expect with deep learning? Can we rely on machines for life critical applications?
Deep learning will certainly out-perform us in few specific tasks with very low error rates. For example, classification of images is task where models can be far accurate than mortals. Consider the case of language translation, today machines are capable of such efficient and economic multi-lingual translation that it wouldn’t just be possible for a person. [Recently MSFT’s speech recognition systems achieved a word error rate of 5.1%on par with humans] While we look into health care where life critical decisions are made, deep learning can be used to improve accuracy, speed and scale in solving problems like screening, tumor segmentation, etc. and not necessarily declaring a person alive or otherwise!
In all the instance we just saw, state-of-art capabilities are developed in very specific and highly verticalized applications. Machine are smarter than us in these applications but nowhere close to our general intelligence in piecing these inputs together to make logical conclusions. From a pure systems and software standpoint, we will need guard rails, i.e. fail-safe heuristics that backup a model when it operates outside the boundaries to keep the fault tolerance at bay.
This is the 4th edition of DVCon in India. What are your expectations from the conference?
While the 20th century is marked by the rise and dominance of the United States, the next 100 years are being dubbed the Asian Century by many prognosticators. No country is driving this tectonic shift more than India with its tech talent. NVIDIA is a world leader in artificial intelligence technologies and is doing significant work to train the next generation of deep learning practitioners. Earlier this year we announced our plans to train 100,000 developers in FY18 in deep-learning skills. We are working across academia and the startup community to conduct trainings in deep learning. I’m keen to understand the enthusiasm of the attendees in these areas and how NVIDIA can provide a bigger platform and bring the AI researchers and scientists community together.
Thank you Vishal!
Join us on Day 1 (Sep 14) of DVCon India 2017 at Leela Palace, Bangalore to attend this keynote and other exciting topics!
Disclaimer: "The postings on this blog are my own and not necessarily reflect the view of Aricent"