On this episode of What That Means, Camille welcomes Ria Cheruvu on the show to discuss deep learning and AI ethics. Ria is the Lead Architect for AI Ethics at Intel's Internet of Things Group. She has a Master’s in Data Science from Harvard University and her research at Intel focuses on artificial security and ethics, uncertain AI robotics, deep learning for the Internet of Things, and computational models of intelligence.

Oh, did we mention that Ria is only 16?

Oh, did we mention that Ria is only 16? This is one convo you don’t want to miss.

The convo covers:

• How AI and DL are related

• Supervised and unsupervised learning

• Generative models

• Adversarial networks

• AI ethics

• GPT3

• Deep learning as it relates to cyber security

• Model extraction

• Red team penetration testing

• Differential privacy

• Model unlearning

• Federated learning

... and more!

Here are some key take-aways:

• Deep learning is a subset of artificial intelligence. There are different ways to use artificial intelligence and one form of deployment is deep learning.

• The goal of deep learning is to parallel or mimic human intelligence so that high-level functions can be performed.

• Deep learning includes multiple algorithms and sub-applications, like reinforcement learning, supervised learning, unsupervised learning, active learning, and more.

• Supervised learning requires that data scientists supervise and correct the computer as it learns. This enables pattern recognition and can be used for everything from cat/dog detection to self-driving cars.

• Unsupervised learning allows for an exploratory analysis of data, without a defined input/output. This type of learning does not rely on human correction. Some applications include: deep fakes, generative art, and 3D objection reconstruction/construction from a single image.

• The major difference between supervised and unsupervised learning is that in the former, the machine is trying to predict an output, whereas with the latter, the machine is trying to create an output.

• Even in unsupervised learning, there is a need for humans to be involved in the process, for ethical and safety reasons.

• Deep learning can be used to improve security threat detection and response by recognizing anomalies and patterns in user behavior.

• Hackers can use deep learning to identify the cyber security defenses of an organization, and also to break the IP of the AI and DL models that the organization is using.

• Ethics and security can be interconnected in some situations and contentious in others.

Some interesting quotes from today’s episode:

“To summarize it, deep learning is our best attempt at trying to mimic human intelligence using algorithms and computational models.”

“In the case of deep fakes, we have issues with AI safety and ethics. Who is going to control the creation of this algorithm? Where are we going to publish its outputs? How do we let the public know that this was generated by an AI algorithm and not by a human?”

“We're starting to build algorithms that don't require as much data, and that can work with small data rather than big data, and still form interesting extrapolations and find interesting patterns.”

“In this cyber security domain, especially, more data is very beneficial for tasks such as malware detection or being able to predict user behavior and anomalies, etc. In these types of situations, there are two different approaches that can leverage deep learning. One is detection and the other is response.”

“The idea here is that just by querying the system and by getting the outputs of the AI model and its confidence scores, you can start to reconstruct that model or even learn more about it.”

“If we can kind of anticipate those problems beforehand using these techniques like red teaming or penetration testing, then I think that that would be a great step forward.”

“I mentioned predictions in retail, which is kind of vague. But the idea is, you’re able to form marketing based on certain data that you're recognizing from your environment.”

"We want to make sure that this data is being encrypted. It's sensitive. It's not being sent to a server, but I'm still providing the user with the insights that they want to know or tailoring the product to their needs.”

“The whole idea with AI security is that we want an end-to-end security solution…All of these techniques could interact with each other, potentially, and are important for securing the entire AI pipeline.”