I am Rama Chellappa, Bloomberg Distinguished Professor at Johns Hopkins University. I am an expert on artificial intelligence, computer vision, and machine learning, and am examining how these areas can be applied in medical fields.

Hey Professor! Just kinda stumbled onto this AMA and as an engineering student I've always been fascinated on the prospects of AI in industry and medicine. But from your experience, what is the biggest hurdle AI is currently facing that prevents it from becoming a complete solution to the problems you are facing? Do you think that this issue will be solved in the near future or is this something that will always be a struggle for researchers and developers?

Modern day AI learns extensively from data. If the data used to train AI systems is not of good quality, the decisions made by the AI systems may not be accurate. In the good old days, AI relied on domain knowledge. We need to integrate domain knowledge and data so that the decisions are more accurate. The other issue I worry about is the vulnerability of AI systems to adversarial attacks. Researchers are actively working on these issues. Like any technology, AI will slowly and surely get better.

Reading this answer makes me wonder why we call it AI at all. Sounds like there is no intelligence present and just machine learning.

Do you worry about how the marketers have (ab)used the term AI?

AI has evolved over the decades. In the early years, AI was driven by domain knowledge. Now AI systems are designed using data.

The deep learning approach that is driving AI these days, basically fits a non-linear hierarchical regression model to data. This approach and new variations are giving very good results on many applications of AI (computer vision, NLP, robotics, etc).

When domain knowledge is integrated into data-driven methods, we can comfortable be in the AI world again!

Dr. Chellappa, thanks for doing this AMA! Where would you recommend an AI newbie start in terms of free trainings, books, or other resources in order to get foundational AI knowledge, focusing particularly on practical applications regarding work, life, etc?

The book by Russell and Norvig is a good one to read. There are many short courses on deep learning, AI and ML that you can benefit from. Youtube has lot of lectures at various levels.

Dr. Chellappa, as a renowned expert in computer vision and artificial intelligence, what exciting developments do you foresee in the integration of AI and computer vision in the next few years? How will this impact the everyday individual?

I think computer vision and AI will positively impact healthcare, personalized medicine, and education. It will also help us to organize our day-to-day activities.

Surgical pathology seems to be a field ripe for automaton with computer vision and machine learning. What is the state of the art, and how soon might I expect to be able to order an automated pathology report?

You are right in that pathology is a natural application area of computer vision, AI and ML. Steady progress is being made in the filed now known as digital pathology. Techniques such as weakly supervised learning are being developed. One issue is domain shift; pathology images generated at different labs may be slightly different; so a method that works on data may not be effective on another data from a different lab. Style transfer techniques are being developed to handle data shift. I think that AI and pathologists will work together for generating more accurate and timely results. This brings the human-AI interaction research into ply.

AI has been quite prominent in society in the past few years in ways that I personally have found quite surprising (e.g. all the AI art stuff). When you were studying AI, did you expect that it would it take the direction it has? What have been some of the bigger surprises in the mainstream application of AI for you?

A great question!

when I studied AI in the late seventies, we were quite happy if AI played games (checkers, chess, etc) well. Although speech recognition was at its infancy in the seventies, we could imagine an AI system that could hold a decent conversation. Data did not play much of a role then. what surprises me is how the data-driven AI contributed to the reemergence of AI. This has to with having powerful computers and sensors that collect data.

Questions : 1. How do you remove or account for bias in your datasets for training models? For medical device/pharma models; how do you ensure bias doesn’t change decision outputs.

1. How easy is getting access to data for training models (how do you address PHI/PII) ? What contracts and derivative use agreements do you typically have with data sources.

2. Any insight into regulatory frameworks for AI use in medical decision making.

For some problems such as face recognition, we have developed methods based on Mutual Information neural estimation (MINE) to estimate bias to gender, age and pose.

We developed adversarial training and knowledge distillation methods to reduce bias.

One can use data that is properly curated and representative of all classes to reduce the effects of bias.

Getting access to data for training models with PII PHI is challenging. New efforts are being made to address is issue. there is a call from the National Inst. of Aging to collect a large data that can be made accessible to researchers who are working on detecting Alzheimers. If health data is available from all over the world for training, we can have better outcomes. I say AI+ world health data = healthy world.

Regarding the third question, FDA has approved nearly 150 AI driven devices and methods. More wll come.

I watched Oppenheimer this weekend, and what struck me was the geopolitics was never the same after the invention of Atomic Bomb.

Would advancements in AI be another such event that can bring a major geopolitical change?

Too early to tell. AI can do lot of good but can also be misused. We need to make AI work for us in good ways.

Hi!

My 4 year old was recently diagnosed with epilepsy. Already, his life has been severely impacted by hospital stays, multiple drugs every day, only to achieve just some minimal control of his seizures (once a week instead of 20+ a day).

I've seen some promising progress in AI being used to diagnose some conditions and create new medications or treatments, but not much progress with epilepsy. Do you see AI helping in the research and treatment of epilepsy in any sort of revolutionary way? Need some spark of hope

Also, curious of your thoughts on other uses for AI for people with medical deficiencies. One of the biggest life obstacles epileptics face is the inability to maintain a driver's license with the diagnosis, therefore not being able to drive, and therefore being at a disadvantage compared to everyone else. It shocks me that there's so little talk about assisted driving being used in a medical sense to 'take over' only in the case of epileptic activity or loss of consciousness. not as an actual replacement for all driving. Are there any companies you know exploring this side of AI as a medical assistant to give people with disabilities the leg up they need?

EDIT:

OH and as an avid vegan, going to need your best vegan-friendly dish I should try cooking.

I am sorry to hear about the recent epilepsy diagnosis for your child. I am not aware of any AI research that is done for detecting/monitoring epilepsy. But researchers are developing algorithms for monitoring seizures in patients who have suffered strokes.

I think your idea of an autonomous car taking over when the drivers are incapacitated is a good one. Since there are many ways in which drivers can be incapacitated, it will be difficult to be responsive to all possible conditions. However, a mechanism by which control can be given to an autonomous car can be included to alleviate the problem you mention.

A good vegan dish is baingan bartha (made of egg plant). Available in most Indian restaurants.

Can you define the term "domain knowledge?" I'm not familiar with this concept.

when I design an autonomous car, I should provide information regarding driving rules, staying in the lane, giving signals when changing lanes etc. Likewise, in medicine, one of the earliest systems known as MYCIN used close to 500 rules related to blood infection.

What are some of your favorite AI tools a person could use in their everyday life?

Google map uses a famous AI search algorithm known as A*. ChatGPT and similar bots are fun to interact with

How far are we away from legitimate androids?

I'm envisioning the day when we no longer have people do dangerous work, body breaking manual labor, and have companionship for lonely people (I'm thinking elderly or special needs, not any neckbeard shit if you catch my drift).

we see the benefit of robots doing the welding operation in assembling cars. We can send robots into disaster areas to look for people etc. Companion robots that can help elderly people are being designed. This will get better as the market becomes more lucrative and self-sustaining.

Who is your favorite cricket player and which dish do you like cooking the most?

My favorite cricket player is Sir Gary Sobers of WI. He was a stylish batsman, spin bowler and a classy criceter.

I also like G.R. Viswanath, V.V.S. Lakshman (great stroke players) and of course Sachin.

I like cooking potato curry.

Hi Professor! As a soon to graduate medical student, what steps/skills can I take to be positioned in developing and progressing the field of medicine as it becomes increasingly interwoven with technology (AI, ML, LLM etc...)?

I think you should develop some knowledge in data science, ML and AI. There are plenty of online courses you can benefit from. Also, reading papers in conferences such as MICCAI and other medical journals that publish papers on AI-based approaches to various clinical problems will be helpful.

Is a random forest classifier appropriate for a dataset with a small number of patients but large number of features? Like 20 patients? Assuming we are getting acceptable results for cross validation. I am performing an experiment involving transfer learning from a small but unique dataset to a larger dataset

The number of patients is too small.

Partial least squares method can also be looked at. Going from a small data data set to a larger dataset is tricky and prone to low-quality inferences.

Greetings professor Chellappa.

I am a medical student. I was wondering how doctors should interact and utilise these new AI tools.

And whether you have any suggestions for them?

Human (Doctors) -AI interaction is an emerging field although human-computer interaction (HCI) is a well-studied CS area. i will refer to the weekly notes from Prof. Ben Schneiderman on human-centered AI.

What makes you an expert considering the technology is so new?

The technology has been there since the late sixties. As you may now, the term AI was coined at the 1956 summer workshop at Dartmouth. I took my first class on AI in spring 1978. As I said before, current AI systems are trained; older systems relied on domain knowledge.

Are there any computer vision problems that you found were too difficult to solve? Any such problems that you've come across recently?

Most computer vision problems are difficult when one compares how well humans do!

Deriving 3D information from a single image has been considered too difficult to solve. We are seeing better performing algorithms for this problem. The most challenging issue is computer vision algorithms lac common sense. They can recognize objects, humans and even interactions among them but do not get the big picture.

How do the machine learning models your organizations employ in Healthcare differ from one's such as chat gets LLM?

What if any, concerns do you have regarding humanity's use of ai and machine learning models?

LLMs such as ChatGPT are still in experimental stage when it comes to healthcare and medicine. My concern is the hallucination aspect of generative AI models. This is like having a friend with some flaws!

I'm aware this is not entirely your subfield, but what are your views on ai allignment/ agi safety?

I am not qualified to comment on ai alignment. Will say that we are not anywhere near achieving agi!

I was intrigued by the hallucinations created in ChatGPT-type models. Do you see this in other areas of AI application? For example, when assessing an image for cancer, does model ever hallucinate a tumor image or something like that?

Generative models can hallucinate. we have a better idea of how often this happens in linear models. The models used in ChatGPT and other large-scale AI systems are so complex, it is difficult to come up good prediction regarding how often and how much hallucinations happen.

Hello Professor,

Non AI person here. I wanted to ask you about academia and a Ph.D. in general. I am just going to start my Ph.D. in Electrical Engineering.

What was the best advice that you received regarding research in your long career in science? and what would be the best advice that you would give to a first-year Ph.D. student like me who is just embarking on this journey?

​

PS. My close friend is a Ph.D. student in your lab :)

Research is something you have to go at it everyday and consumed by it. Unlike BS and MS where you are expected to do well in courses, in research you have to create something novel. Passion, hardwork and some amount of luck (being at the right place at the right time) all play a role. Passion helps when things do not go as well as you like in research.

Given the big question of security and ethics with platforms like chatGPT etc, can we trust AI with medical data? And how can we be sure an AI won’t be abused? For example with racial discrimination in the base code.

While AI can do good things, it can also be misused by (humans!). We need to constantly monitor how an AI system functions and provide remedies when bias, vulnerability to adversarial attacks etc occur. with ChatGPT, hallucination is a big worry. I say, Trust but Verify (President Regan's words!)

Some inputs and feedback on how to build platforms which include AI etc...? Platform thinking? Thanks

You can use PyTorch to build AI systems. Data-driven AI is enabling end-to-end AI platforms. One should have interfaces that link AI software and sensors.

How can AI and Data Analytics be applied to the energy storage space? Materials space? Thanks

AI is being used to discover new materials Check out HEMI at JHU.

AI can be used to create efficient energy storage protocols.

Hi Professor Chellappa!

As an ML researcher, I am sometimes concerned by the excessive focus placed on LLMs in the past year, even at the expense on other research areas. What is your take on this? Do you think this state of affairs will prevail in the coming years?

ChatGPT and similar bots have received lot of attention. But there are many other AI research activities are being pursued in public health, medicine, smart transportation, etc which will have long term impact. Its usually the case, that there are many competing research areas, some more glamorous than others. Eventually, research that leads to tangible benefits will thrive.

Hi Professor

Would it be possible to create AI that can predict future AI trends? Like some sort of market prediction bot but for 'AI Trends'? Do give me credit if a paper comes out of it. lol XD

As Yogi Berra said, it is difficult to make predictions, especially about the future!

AI as it is built these days works best with data collected in the past and can predict to some extent the future. Not sure how effective it will be for predicting trends in AI.

Very few of us would have predicted the reemergence of deep learning (which is driving much of AI these days) in 2011!

Important work. I wish Kaiser could let the AI loose on my medical data and see what underlying causes they might suspect and tests they might suggest.

For smart transportation, could you improve busy traffic lights by having the AI watch oncoming traffic and adjust the light timing to favor turning red right before large gaps rather than fixed frequencies? In short, have them do what a human would do when directing traffic.

Yes to question 2.

JHU has PMAP which can do analyze medical data and look for underlying causes. Check it out.

Will A.I. be mad at us for not fixing the climate?

We should be at ourselves for not fixing the climate! Just got back from a trip to Greece.

Was in Rhodes. Sad to see the fires in Rhodes. We are heating up the planet!