PODCAST: Making AI Real
Episode 5: Debunking AI Myths
Listening time: 15 minutes
Debunking AI Myths
In this episode of the AI to Impact podcast, Divyansh Mishra and Anil Prasad from AI Labs, BRIDGEi2i discuss the myriad misconceptions of artificial intelligence and decode the real deal with AI. They talk about the buzzword that used to be AI and elaborate on the science behind Deep Neural Networks. Tune in to the podcast to know the differences between General AI and Applied AI, the science behind the black box that is AI, and whether AI is a function of datasets or a programmer’s intelligence. Listen Now!
In software engineering, there is an adage, ‘Software is only as intelligent as the programmer who programmed it. That’s true for today’s AI systems. There are techniques like online learning, but they too are subject to programmer ingenuity and intuition. You simply can’t build solutions that you don’t understand.
Divyansh: You are listening to AI to Impact by BRIDGEi2i, a podcast on AI for the Digital Enterprise, wherein we gather thoughts and opinions of market leaders and industry experts in the domain of AI and analytics, and walk you through the evolution of AI, how it’s changing its course over time, and how we see a real-time impact on varied industries, and how it has been transforming processes and functions at organizations.
Now, since the wave of AI is pretty dense, and people are jumping on the AI bandwagon with too much curiosity and enthusiasm, one cannot emphasize enough on how AI has become a part of our everyday life. And while all of this happens through machines, there are misconceptions, and the fear of adoption around AI is definitely undeniable. And so to speak, there are so many conflicting messages around what AI is, what it is not, what it can do, and what it cannot. So today, we’ll try and break some of these misconceptions around AI.
So welcome to our podcast series Beyond Theory with AI Labs, and I’m your host, Divyansh. And my guest today is indeed a very special person too. In fact, I’ve started my professional career under his mentorship. His journey to the field of AI is an interesting one. He has over a decade of experience in the field of AI, which he started with his paper on machine learning in 2009.
After his postgraduation from IISc in computer vision, he had an entrepreneurial stint co-founding a company that was into video analytics. Today, he’s leading some of the crucial AI projects at BRIDGEi2i. And at AI labs, he is involved in building innovative AI and ML capabilities. He has published over eight papers on computer vision. And I still have to come to terms with the fact that he had published his first paper on machine learning when I was in the seventh standard.
Anil: Hey, that makes me feel older than I am, Divyansh. But thanks a lot for that introduction. It’s great to talk to you on this forum. And it’s great to see your career progression, blooming into a data scientist.
Divyansh: Thank you so much, Anil. And it’s a pleasure to have you. So, let’s get started. So, I remember when I was studying, and there were a lot of buzzwords like big data, Hadoop making the rounds and then came data science. And now we feel everyone is so excited about AI, ML and NLP. So, as professionals in the analytics industry, we ourselves have observed a huge mind shift in organizations in recent times. And this brings us to the first myth about AI, that the field of AI is pretty new. So, what are your thoughts on this?
Anil: AI has been around for a long time. It has a 70-years history, starting 1950s. It’s been a bumpy journey for AI research with cycles of enthusiasm when tall and fancy claims were made and a period of disinterest when these claims did not come true. But to give you a sense of how bad it had gotten when I was a post-grad student, it’s not long back; it’s in the 2010s. Very few labs called themselves AI labs, even if they were exclusively working on AI. So funding became hard to come by as it was considered pseudoscience. While today it’s the buzzword that sells.
Divyansh: Got it, Anil. It definitely is a buzzword that sells. But I keep hearing that AI would be able to solve any kind of problems, and so when we talk to our clients, they generally describe their problems and tend to imagine solving them using the help of AI. So, is AI really the solution to every problem around us?
Anil: That’s a big one and also a bit nuanced. Only certain problems can be solved with AI today. AI today is largely powered by an algorithm called a deep neural network, which is a type of machine learning algorithm. It’s also called software 2.0 because, unlike traditional coding, the system automatically learns to solve a problem by using huge amounts of data. This is very similar to a show and tell game we play with kids that teaches them to differentiate between, say, a cat and a dog.
Divyansh: Absolutely. And if this is a show and tell game, how would a network differentiate between features of these animals?
Anil: Deep learning systems are essentially large networks with many layers constituting an artificial neuron that fires when a certain set of its input neurons fire. This set of inputs can be thought of as hidden patterns, and the neuron firing in response to such stimulus is very similar to the human brain, and hence the analogy with the neurons.
Divyansh: And once the network has differentiated these set of features, how would a network learn to do this process?
Anil: To train such a network, we rely on something called backpropagation, where the neurons that contributed to a wrong prediction will be modified proportionally to their contribution to the wrong output. And we are tweaking them over time so that they become better at that task by a small amount, but over time, they become better able to classify or any other task. So, let’s say as a neuron starts detecting the eye of a dog, while another the nose of the dog, a neuron at a higher level, will look at the output of all these neurons and fire whenever the right combination of the dog faces present. Such relational hierarchy is itself a product of the learning algorithm and not hardcoded. So, for the problems that neural deep neural networks can solve, huge volumes of data are required. We need to show the system multiple annotated images of dogs before it makes sense of the association and learns the underlying concepts. And data becomes crucial and as important as the algorithms to get the job done.
The good news is that networks trained on one domain can be used to solve problems in some other domain. That says that they’re re-usable. Let’s say to detect other animals, and the dog eye might be a very good hidden path, parameter or pattern since eyes are universal. Hence, we don’t need to train a special AI system from scratch. But we will use one that’s already built for, let’s say, the cat and dog example. Further, these problems should have underlying patterns. Though today, there are some new algorithms that are emerging that tackle these more complex problems. But they’re not still mature in terms of commercial level. This will lead to additional resources and tools in the toolbox of AI practitioners when they become available. This is similar to traditional methods, which often perform better with limited data. We routinely use the solution to bootstrap an AI system over time.
Divyansh: Yes, this is a very fast-changing field, where progress is being made at lightning speed. And this probably proceeds to the next myth about AI that AI function is a pure function of data, or is it about the intelligence of the programmer who has designed the solution?
Anil: In software engineering, there is an adage, ‘Software is only as intelligent as the programmer who programmed it.’ That’s true for today’s AI systems, too. There are techniques like online learning, but they too are subject to programmer ingenuity and intuition. We simply can’t build solutions that we don’t understand.
Divyansh: And yes, as humans, we have intuitions and expect the robot to learn and think like us. But are these systems working in only one domain?
Anil: Yes, there are rules built on top of an applied AI which are used to solve only certain problems. Today, we can best over a machine with the ability to understand images, have limited cognition of videos, read text, comprehend spoken sentences, speech, and limited reasoning skills.
Divyansh: So, Anil, what would be the difference between applied AI and generally,
Anil: Today’s AI are what are referred to as applied AI systems that simulate intelligence without necessarily having solved all of the underlying complex reasoning. In contrast, General AI is the ability of machines to reason and think like humans do. And that’s not possible today. Applied AI looks like there is some logic or rationality that we bring to the table. For instance, we might say that there is a sun and then the lighting, and we can solve for the entire scene in a three-dimensional nature, but a neural network would be able to map the input and output in its own set of ways. There might not be any reasoning behind how it comes to the conclusion of what the cat is and what a dog is. We are reaching towards such an AI singularity; nobody knows if and when that will be achieved. We know that it will require a different type of algorithm, which is not yet known.
Divyansh: So, I cannot order a smart robot anytime soon?
Anil: No, unfortunately not yet. There are newer fields like reinforcement learning, GTP like unsupervised or semi-supervised methods, the true promising results and have different trade-offs and applicability to certain problem statements. The distinction between Applied and General AI will keep reducing, giving us better and better illusions of General AI, will continue to build solutions that are feasible considering a large number of constraints such as competition or power, business value, fit so on, but none that is actually a general AI for now. But for today, when we say AI, we mostly referred to deep neural network solutions.
Divyansh: And yes, people think AI is a black box, and organizations do think AI will magically solve all of their business problems. But does nobody really understand how deep learning actually works?
Anil: It’s perfectly fine for organizations to think that AI is a black box and use it innovatively in their business process. But deep neural networks are complex systems, which are hard to understand. We would need to understand infinite-dimensional spaces and other exotic maths to master the learning mechanism we talked about earlier. So, you know, it’s easier not to explain it to people. We do understand how deep learning systems work, but explaining it is much harder. Folks spend years in post-grad developing an intuition for the subject. Yet certain behaviours might be unexpected and counterintuitive. To further complicate matters, it’s still data science, and data sets have a mind of their own, with no standard way of solving them.
Divyansh: Then, if that is the case, do we really build specialized solutions to problems and park the general artificial intelligence for now?
Anil: It’s a balance. We need to make special modules, but we like solutions to be as generic as possible. We have come a long way in terms of the generality of the solutions. Today we have a plethora of robust neural network architectures. We exploit transfer learning with pre-trained networks from one domain to the other, like we mentioned that we had sought sophisticated cloud support with solutions like auto-ML that accelerate this process, sometimes smart pre and post-processing hacks and reformulation of the problem leads to amazing results. And also, why solve the harder problem?
Divyansh: Oh, okay, but what do we mean by a harder problem here?
Anil: We know that General AI is not part of the toolkit an AI practitioner today has, but that doesn’t stop us from doing interesting things with it. Companies influenced by jargons like AI would want an AI solution today. Yet if the problem requires different solutions, we give them Applied AI solutions that have limitations like power compute. Actually, cost computations might make them unviable and diminish ROI. So, we end up using simpler solutions that have better characteristics. Further solutions today are heterogeneous. What I mean by that is, with only some AI components, today, applied AI is like the engine of the car, but we need other parts as well. Tomorrow, we can swap the engine with, let’s say, General AI whenever it becomes available, right? Having digital transformation, optimized workflows, and AI solutions in place allow for rapid adoption of these emerging trends. And at BRIDGEi2i, we have been part of such digital transformation for a number of businesses.
Divyansh: Absolutely, Anil. So we have got clarity on so many myths and folklore around AI. Number one, businesses need to have a realistic and accurate view of what AI can do and cannot do. Number two, the AI hype is not new. AI has been there for a very long time now. Number three, human intuition and intelligence can never be replaced by AI systems because AI is not a one-stop solution for all business problems. It is not a plug and play solution. It is definitely not a magic wand. However, it sits on top of our existing systems. And as Anil very rightly mentioned, AI is just another tool in our toolbox. That brings us to the end of this very interesting conversation with Anil, Manager at AI labs at BRIDGEi2i. Any parting thoughts for us, Anil?
Anil: There’s a lot of research going on around the globe to explore other solutions that could see sophisticated algorithms emerge. They’ve achieved a flashpoint where deep neural networks have achieved maturity, being integrated and transforming businesses, and are doing unbelievable sci-fi stuff. There’s a lot to learn from the previous history of AI. It’s important that we understand applied AI for what it is and avoid the pitfalls, the hype and the disillusionment. Hence, it’s important to beat the myths around AI.
Divyansh: Well said, Anil, and thank you so much for joining us on this podcast. It was a pleasure to have you with us today.
Anil: It was a pleasure to interact with you, Divyansh. Thanks a lot.
Divyansh: To all the listeners. Thank you so much for joining us on this episode of the AI to Impact podcast. To subscribe to our podcast and let us know in the comment section if you found this interesting. We will see you the next time. Until then, goodbye!
With new data formats, evolving data ecosystems and AI technologies becoming more mainstream, the intersection of Data Science and AI-powered interaction systems present a whole new gamut of opportunities for digital enterprises. In our latest AI to Impact podcast series, Making AI Real, we interact with several thought leaders, BRIDGEi2i business heads, domain experts, and reputed AI and analytics leaders to learn more about the ground realities that further AI innovations. Stay tuned to know more!
Anil Prasad MN, Manager – AI & Innovation Lab – BRIDGEi2i
Anil Prasad MN is a futurist and a robotics enthusiast, focussed on cutting-edge research to solve real-world problems at scale to better humanity. He has a masters degree (MS – Research) specialising in Computer Vision and Artificial Intelligence from the Indian Institute of Science (IISc), Bangalore. He has published multiple research papers in leading ACM and IEEE conferences. Anil has worked in various startups and led complex technical projects successfully, impacting all aspects of the project lifecycle. He currently works in the AI Labs at BRIDGEi2i. Psychology, neuroscience, metacognition, the process of learning to learn, and the impact of singularity and the emergence of quantum consciousness are some of his key interests. Reach out to him for a stimulating discussion on these subjects.