- In this episode
Welcome to our first episode of No Bias, where we discuss different perspectives on the emerging and ever-shifting terrain of artificial intelligence and machine learning. In future episodes we’ll dive deeper into the nuts and bolts of developing and training models, philosophical issues, and existential concerns. But since this is our first episode we decided to begin with the basics: AI versus ML. We offer definitions and historical background of how they have evolved over the past few decades into the current state. And then we will peek around the curtain to discuss the future of the industry.
Melody: Hey! Welcome to our first episode of No Bias. Where we discuss different perspectives on the emerging and ever-shifting terrain of artificial intelligence and machine learning. I’m your host Melody Travers and I’m here with machine learning researchers Nikhil Kumar and Saurabh Bagalkar.
Saurabh: Hey Melody! How’s it going?
Nikhil: Hello Melody! How are you?
Melody: I’m good! Thank you guys so much for joining me today. Nikhil, Saurabh, and I all worked together and have enjoyed chatting about the latest developments in the world of AI. We had so much fun, in fact, that we thought why not make a podcast and share our musings with the world. So here we are. Each episode, I get to pick their very big brains and hear their different ideas and perspectives on the frontier of AI technology. Saurabh, your background is in Biomedical engineering, and Nikhil yours is in Statistics. So, can you tell me a little bit about how your perspectives differ, and how that informs your research?
Saurabh: Sure! So biomedical engineering has various specializations within it, one being medical imaging which is very closely related to computer vision, and in recent years deep learning and computer vision have been pretty much hand in hand because the research in both of them have been very, very close-knit.
Melody: Awesome! How about you Nikhil?
Nikhil: Right! As you mentioned my background is in Statistics, and I initially started out sort of focusing on the math and statistics in machine learning, and then once I left college and started working, I basically was involved in applying statistical models and machine learning models to businesses and then kind of solving business problems using machine learning.
Melody: Right on! So in future episodes, we’ll dive deeper into the latest topics and philosophical issues, but since this is our first episode, I thought we would begin with the basics: AI versus ML. Andrew Moore describes artificial intelligence as quote “The science and engineering of making computers behave in ways that until recently, we thought required human intelligence.” How would you define AI?
Saurabh: Yes, so that’s very interesting quote and actually yes. AI attempts to mimic certain aspects of human intelligence so that we can automate human tasks, and it is a very broad term that has a lot of facets to it.
Nikhil: Right! So as Saurabh mentioned about automating human tasks this really began in the 1950s, when computer scientists and cognitive psychologists met with the common goal of trying to understand and ultimately programmatically mimic human cognition. And then with that work that inspired other researchers to follow them. However, they faced some difficulties namely with the fact that the work they were developing did not seem to integrate well with the current technology stacks and methodologies that were used at that time. This caused a bit of an AI winter as less applications were built because nothing was integrating. As of late 1990s things started to change. The advent of personal computing and the internet became more ubiquitous and improved and advanced hardware, optimized software, as well as the data storage and retrieval became much more common and this allowed AI research to further develop. And then at the time the main applications were in the area of self-playing games, things like chess or tic-tac-toe or snakes and ladders and those were sort of the first set of successful use cases for AI.
Melody: That’s right! I remember back in I think it was ‘96, ‘97 when Deep Blue won the championship, the world class championship in chess against Garry Kasparov. That was huge news at that time.
Saurabh: That was phenomenal, yeah.
Melody: Yeah! And then it was I think last December, when AlphaStar beat the first professional Starcraft player and we see this huge shift right from a sort of static game to one that’s live action. So where are we today?
Saurabh: AI has grown by leaps and bounds in the last couple of decades. It has infiltrated almost every field. Nikhil gave a very interesting history of the evolution of AI. Right now, AI is very heavily fueled by access to large datasets, an amazingly huge amount of computational power, hardware resources, and specific domain knowledge. So currently AI has developed into its own entire industry, which is solving sophisticated problems in almost every industry ranging from self-driving cars to medical diagnostics, stock market trading, and advanced gaming.
Melody: Wow! So as researchers you guys are on the cutting edge. Can you give us a glimpse into where this technology is going in the future?
Nikhil: I think the key thing will be the fact that AI will be increasingly ubiquitous in our lives. So as Saurabh mentioned, all these different and interesting use cases, self-driving cars and medical diagnoses, these are things that will affect our lives on a daily basis. And so there’s going to be all sorts of new and exciting innovations and new discoveries from all of that which is all great. I think on the other side, there’s gonna be some unanswered questions in terms of how the regulatory side and how public policy keep up with these AI advancements. And I think that one of the coming challenges of the next couple of decades is to answer how does society properly manage the new innovations coming from AI.
Melody: So what you’re saying is we still need some liberal art students as well as the STEM students?
Melody: Cool! That’s good to know. That’s my background so I’m glad we’re not totally obsolete yet. So to recap, AI attempts to automate certain tasks by mimicking human intelligence. Is that right?
Saurabh: That is right. Yes.
Melody: Awesome! Okay so let’s move on to machine learning. How is it distinct?
Saurabh: Sure! And that’s a good question. This deep blue, deep mind, deep learning are deeply misunderstood because machine learning and AI, although they are very much related they have distinct boundaries. So machine learning is one of the common ways to achieve artificial intelligence. It is a specialized subset of artificial intelligence which involves learning from the past experiences which are stored in the form of data. Machine learning improves its learning with minimal human intervention if any and that’s how I see it.
Melody: So artificial intelligence is sort of an umbrella field and then there’s machine learning as a subdiscipline and then even further along there’s sub-branches of that?
Nikhil: Exactly! And as Saurabh mentioned, the key characteristic of machine learning is that you’re learning from data. That’s what the “learning” refers to. And with that there’s many different ways to process data, to learn from data, to analyze data. There are about four major subdisciplines. So historically the two common ones were supervised learning and unsupervised learning. Supervised learning is all about task-driven analysis from labeled data so this is the type of data that would have a dependent variable or a target variable and you’re trying to solve a very prescriptive problem.
On the other side, you have unsupervised learning which is a data-driven form of analysis and there’s no labels, it’s unlabeled data. The idea here is to learn from data in a generalized way. So things like high level characteristics, aggregations or groups patterns that type of learning is what would call unsupervised learning.
As of recently there’s also been two kind of newer but equally as exciting areas of machine learning called reinforcement learning and deep learning.
Reinforcement learning is the idea of learning from past mistakes. The idea that a machine or some algorithm makes a decision and there’s a cost and a reward associated with each decision and that combination of costs and rewards are optimized until a truly ideal policy is discovered.
And then even kind of more modern is deep learning. This is a very exciting and very new area of machine learning. The idea here is that this is a very specialized area of supervised learning in which a specific type of model is made called a neural network and these are designed to learn from very, very large sets of data and with the specific focus on learning the nuances of very large, very complicated data sets.
Melody: Cool! Okay, so can we go back a little bit into a historical context. You gave a little bit of an idea of how it’s evolving, but how far back does machine learning even go and how did we get here?
Saurabh: Yeah, that's an interesting question. So, machine learning was obviously not called machine learning 20, 30, 40 years back, because at its core it is statistics, probability, and math applied to smarter algorithms. So, I think a group of researchers and cognitive scientists came together to apply some statistical modeling which included regression, clustering, vision modeling techniques to smaller data sets and this basically evolved in 80s and 90s with better computational power, allowing for using the smarter algorithms on larger datasets, right?
During the 2000s, we had the advent of big data by Google and Facebook, which produce petabytes of data every day. This led to smarter machine learning algorithms to be developed to analyze all those data, say for example on social media or internet searches. This in turn sparked widespread interest basically among the researchers and businesses alike to adopt machine learning as part of their overall strategy.
Melody: So you mentioned that companies are starting to incorporate machine learning into their strategies. How are they doing that these days?
Nikhil: Right! So given where we are currently with machine learning based on the time frame Saurabh mentioned, it is an increasingly integral part of any business of any industry of their overall strategy. It’s becoming almost a necessity to have some machine learning strategy.
Currently there is a great deal of interest in developing deep learning algorithms. So as I alluded to deep learning is a new, emerging, and exciting area of machine learning. As Saurabh mentioned, we're now looking at data that is on the scale of petabytes. Deep learning is something that really will be the tool that when done properly or at least has the potential to be the tool that will process that data, provide very valuable insights, recommendations or analysis and does so in a computationally responsible amount of time.
Parallel to that, there's also, all sorts of hardware and software improvement things like GPUs and TPUs. All of this is going back to the fact that with deep learning we have convolutional neural networks that are the modeling tool that will unlock these solutions. There's also a great deal of media attention towards machine learning and this is both by actual media people as well as by investors, venture capital firms, and things of that nature. We're seeing a great deal of excitement and just the ability to invest, to gather capital and invest in machine learning AI startups that are able to develop this sort of applications.
Melody: I recently read an article in Forbes that talked about this new wave of technology of AI and machine learning as spurring on a 4th Industrial Revolution or Industry 4.0 and they said that its disrupting as you guys said, every industry in every country, and at an unprecedented speed. Peering behind the curtain a little bit, how do you guys see this developing in the next five years and maybe even further out into the future?
Saurabh: So Nikhil gives a very interesting insight into the machine learning aspect and what the companies are doing and that will continue to evolve and grow. You have a very interesting question. It's like asking what is the future of information technology in the 90s because before the 90s, information technology was like people did not know what it is capable of, what can it do? But right now that is state of machine learning we have a glimpse into the future right now, but not the complete picture of what it is capable of, but it is continue to evolve at a rapid pace. Right now at the current state, it is a very important part of our daily lives. As I mentioned, when we use our smartphones or when we use some other high specialized gizmos, we use artificial intelligence in some way. But in the future if we are speaking about 10, 20, 30 years down the line self-driving cars and all that can really affect our daily lives. That's the pros of the technology some of the cons of the technology can be an example of deep fake, which is basically a fake video of some famous personalities that can influence the masses and that can create havoc. There are downsides to technology as every technology has some positives and negatives, so AI has some negatives to it, but it will be interesting to see what the future beholds.
Melody: Sure, and what are some factors that are holding us up or impeding this progress?
Nikhil: This is all really about data in many ways. Saurabh and I both mentioned a number of amazing technologies and algorithms that are creating this great output but none of that is useful if you have bad data going into the model. I don't think we can overstate the necessity to have a good cleaned properly labeled training data sets. When people talk about how data is the new oil. This really goes to that metaphor because there's a process to refine oil. Similarly there is a process to refine data and you can't just use raw bad data or you can't just use crude oil in a manufacturing process for example.
Melody: So, I've heard it talked about as garbage in, garbage out
Nikhil: Exactly! Yeah.
Melody: What does that mean?
Nikhil: I think it means, basically if you have a bad data sets, that’s the “garbage in”, then basically the output will be garbage. If we talk about, for example, a self-driving car if it's not trained on a good set of traffic data, then it won't know what to do if there's a lot of traffic and it’ll make bad decisions. This is bad for a number of reasons possibly including the fact that you have fatalities for that, right? So having well curated data sets is very, very important and that's something we can’t overstate. As we also alluded to in the past, there's a big regulatory aspect to this and there are a number of unanswered questions. I think the need to sort of bridge the public policy side with the machine learning side is going to be very important.
Saurabh: Nikhil mentioned a very interesting metaphor regarding how AI is the new oil, and to take that metaphor a bit further, a very interesting article in MIT Technology Review states training that a single AI model can emit as much carbon as five cars in its lifetime, so that's something of a concern for environmentalists in the future.
Melody: Oh man! Well I definitely don't want to end this conversation on that note, so can you guys give me a takeaway example of how this technology is going to change the world in a really positive way?
Nikhil: Yeah, I think there's a number of examples of that. A really good example stems from the work done by the Bill and Melinda Gates Foundation. They're using machine learning to better improve and optimize agriculture in poor countries. For example, crop farmers in poor parts of Africa can now scan an image or a seed on their phone or on some applications and they'll know what the yield of that plant will be and they can better optimize how they farm. This directly affects how they get income, their livelihood, how they support their families, and all of that. That is an example and underneath the hood of that is deep learning and how that will really help in a very positive way of literally lifting people out of poverty because of machine learning.
Saurabh: And they are using actually, as Nikhil mentioned, computer vision to basically take the images of the flowers on the crop and to understand those images so kind of like image recognition and using that image recognition to have you know, for the betterment of mankind in that area of the world, so that's a very positive impact for work that deep learning and artificial intelligence is doing overall.
Melody: So, we're sharing the hive mind, I love it! Well, thank you guys so much for coming in and talking to me about this today.
Saurabh: Thank you, Melody.
Melody: Come back next time for our discussion of the ways in which data is and is not the new oil. Thank you so much for listening, this episode was produced by Melody Travers. Big thanks to our experts Nikhil Kumar and Saraubh Bagalkar. Music by Melody Chebrellan and sponsored by Alegion - “Integrating human and machine intelligence to produce ground truth data for enterprise, AI and ML projects at any scale.” For more information, check out our website Alegion.com and don't forget to tell your friends about us and subscribe. Thanks!