Good morning, everyone. I am Chirag. I work at Nutanix in cyber security with AI ML. so today I am going to cover a presentation on pioneering AI frontiers using deep learning based techniques. So what is reinforcement learning? So reinforcement learning right now is a very new frontier in deep learning, and, Reinforcement learning is, a learning which involves complex decision making and there are a lot of reinforcement learning based models, which involves neural networks, which are currently in use. And the current market of, AI is going to increase. As you see over here. It'll increase by one 90 mil billion by 2025. Maybe by 2030 it'll be. Almost 10 times more. And these kinds of deep learning techniques currently can outperform traditional ML models. So that is why deep learning based techniques are a lot in demand. These are, used in a lot of applications like image classification, autonomous driving, and, many other, techniques. So the main core over here is a neural network and reinforcement learning basically is a technique which involves interacting with environments and using trial and error based approaches. So wherever the neural network works, it tries to find the errors, then it will try to minimize the errors and in turn try to So that is how the algorithm works and combines deep neural networks and also it has state space management. Now we come to the milestones. The AlphaGo's victory over Lee Sedol highlighted superhuman strategic planning. Then in 2018, we saw UC Berkeley, which enabled robotic kind of grasping, high success rate. And in 2023, Waymo self driving cars, they achieved 35 billion simulated and 20 million real time miles. So these are very good achievements, which have reinforcement learning using neural networks, which are being applied. Now we come to. Accelerating training using transfer learning. So what is transfer learning in reinforcement learning enables agents to apply knowledge from one task to another. And significantly, this reduces the. Timing, training time, and effort. So this is a key advantage of this algorithm. It has faster adaptation to new environments. So these kinds of neural networks in reinforcement learning have a capability to adapt to any kind of conditions. And reduce training data requirements. So machine learning algorithms work by training the data and then generating some of the results. So the more you train the models, there is a chance of overfitting and that kind of scenario. So neural network over here is very. Optimized and it will try to train the model with reduced data Requirements and it will try to provide generation and output in a short span of time with high accuracy So that is why these kind of applications like autonomous driving, robotics, then game AI, all these kind of applications, they are using reinforcement learning and, these kind of, machines, they leverage, such kind of neural networks, which will take some of the training data and generate proper, kind of output, which is capable of, Of improving the experience. So for example, in autonomous driving, the neural networks are so advanced that they will be able to detect anything in the road, any issues which are happening in the road, and it will apply the brakes, then it will accelerate at certain point of times. And similarly, in robotics, the robots can detect any kind of boundaries or any. Places where they are not supposed to go and games, they perform quite better and it will help in better game generation and using less training of data. So this is where reinforcement learning is a boon. And, then we come up over here in this slide, it, it provides information about the generative AI market. So how much is the market capitalization right now? And by 2030 maybe the project, projected, capitalization will be a lot more, almost around 33.7%. So it would be, I guess it would be like, around 40 billion. 30 to 40 billion. Yeah. And, over here, we also have GANs, Generative Adversarial Networks, and VAEs, which is Variational Autoencoders. These are two different kinds of models, so Generative Adversarial Networks, they focus on generating data and through training, whereas Variational Autoencoders, they are actually, they create compact. certain kind of data representation for generation, some kind of analysis, that kind of thing. So these two are very different kind of models. And let's go over here in this slide, how they work. So GANs are two networks, generator, discriminator, they compete to improve data generation quality. And basically it will be utilized to generate certain kind of things like for example, image generation. Then we have video synthesis. Then we have medical data augmentation. So which involves classification and then we have drug discovery. So all these applications are it's covering something and using model training to perform certain kind of generation, image generation, video synthesis, then medically in medical field, it will try to improve the classification. which is a very important thing. Like you, Using certain kinds of inputs, it will generate certain kind of patterns and we'll try to find out the pattern and classify into better thing or classify a medicine or classify certain kind of things. And similarly in drug discovery, GANs will try to kind of classify and generate certain kind of modern molecules for medicine development, maybe fighting certain kinds of illnesses. So we see GANs. will have a lot of capability in next few years where it will not only improve the drug discovery, but also help in image generation and maybe in cyber security or security and that kind of applications, it will have a very huge impact. Now, we come to VAEs and over here how VAEs work, they encode the data into latent space and reconstruct it, and that is how it will learn, and it will become more efficient. one good example over here is anomaly detection. so whenever certain kinds of neural network, they find certain kinds of issues in the data, it'll classify it as an anomaly. anomaly detection. Then we have dimensionality reduction. So over here it'll have compressed gene data by 98% and it'll, reduce the dimensionality. And thirdly, another use case would be music generation. generated, 16 bar melodies and it will have a lot of, satisfaction rates. So these, these don't do any kind of generation, but they try to do certain kinds of, anomaly detection and that kind of thing. And these. techniques will become popular in next few years, and it will be utilized in a lot of other applications. Now we come to AI in cyber security. So in this cyber security, it will do a lot of things like threat detection, threat monitoring, and vulnerability assessment and fraud detection. We do see a lot of, in, cyber security, we see a lot of, certain kinds of threats are not monitored properly and it leads to a lot of, issues and, a lot of threats and a lot of vulnerabilities. And there are, Current tools, which are available in cybersecurity, they don't detect such kind of vulnerabilities and threats. maybe in next few years, AI using GAN and VIEs with, these techniques, they will improve the threat detection and anomaly detection And basically. in cybersecurity, most of the companies won't use these kinds of tools and they will rely heavily in this machine learning models and such kind of things to improve their security posture and, use AI based threat detection. It can also do maybe in future AI based techniques. We can also do threat modeling and, identifying risks, within the organization and maybe. Possibly detecting DDoS attacks and such kind of things. Now we come to future trends in AI. So in future trends of ai, we have graph neural networks, which can enhance, AI's capability to process graph structured data like social networks, recommendation systems and drug discovery. So graph neural networks are also very important kind of neural networks, which can be utilized in recommendation. for example, we see in Netflix. and Amazon prime. There are a lot of recommendations and certain things which are being done. So these company maybe in future can leverage a graph neural networks and such kind of sophisticated algorithms to improve the accuracy of recommendation. Then try to process text image or video and certain kinds of applications. Then we have AI in IoT and certain, other applications where AI can have a huge impact. this is what we have seen in enforcement learning. It has a lot of unparalleled capabilities in doing complex tasks like autonomous driving, robotics, industrial automation, and it uses GANs and VAIs to unlock a lot of new possibilities. Helping creativity and problem solving these kind of approaches can help the organization to improve upon their security posture, improve the drug discovery and have a lot of big impact in decision making capabilities. Even in cyber security, maybe 30 which are being used, they might be either, GANs or VAEs, which can be used to detect certain kinds of threats and improve the security posture within the security org. So that's it from my side. If you have any questions, please let us know.