Bhat's Notes

This topic of Future of AI is very close to my heart. There is a clear divide. One side believes AI is overhyped and AI can never match human intelligence. The other side believes that AI is a threat and it will soon match human intelligence and rule the world. I like this topic so much that I wrote a story about it by sampling a day from 2040.

For me, future of AI truly lies in the hands of AI researchers of the world including the ones who attended summer school at IIIT. Here is an interesting panel discussion on the future of AI by researchers from IIIT, Inria, IBM Research and a seasoned entrepreneur.

PJ Narayan: A decade ago, as researchers we alienated ourselves from AI. Now we are proud to say we are in AI :). Things can only go forward. Now a days, startups are adding .AI to their names whether they really use AI or not.

Karteek: As we speak ICML conference is happening in Sweden. I don’t think any robot in the world that can do something as simple as walking from that far end of the door into this room and out of this room from the left most door here. It is all media. AI may take over monotonous jobs but this is nothing new. Each revolution has taken over many jobs and probably created much more. You might have heard of self driving cars. If unfortunately there is an accident, who is responsible? The maker? The researcher? The driver? Who? Legal aspects have to be worked out.

Safer cars. AI radiology. AI translators. Many applications are feasible. many things that we don’t know anything about how to model such as common sense. How do you define common sense mathematically? If we can’t do it with math, how could a machine do? How about employing a robot to do daily house chores? We don’t know how to do it. Too many variables and unknowns. What if a self driving car that has never seen a plastic bag flying in the air. You and I go through but AI car has never seen it before. This is a edge case. We shall worry about what real experts who work in AI say, and not what leaders who aren’t directly involved say about it. Selective information gets propagated in media. We shall look at all these things as we move forward.

Vikas: Every day there is a new Arxiv paper. What are the new areas to focus on? AI soon will be democratized.  I would never write a sorting algorithm and we just take a pipe. I see this happening with CNNs and LSTMs. I am not sure how many of you have implemented LSTM from the scratch. We would not bother about implementations and we work on Caffe or Tensorflow without having to go through the nitty gritties. There is a little bit of sad story as we try to fine tune the hyper parameters. Pick an library, collect data set and find the best hyper parameter that works. Rather than calling ourselves AI scientists, let me call myself hyper parameter tuners. None of the models will be 100% accurate. Take ImageNet. Talk not Top 5 but talk Top 1. Then it is probably 85%. If the classifier is not accurate, I can bring the human in the loop. Robustness is the key. Till some of the systems are provably robust that they can work on different conditions, we can’t get AI to do things well.

IBM Research got three pillars.

1) Interpretability: Most interpretable models are probably linear or logistics models. We can show the importance of a variable to a doctor and convince him. Take visual search. Can you explain why the search results have that order and explain why an image took 6th place? Can we? Lot of newer researches are emerging. male the algorithm more explainable. By design or even by interpretation. CLients would rather talk an average system that can explain better than better system with better results but no explanation.

2) Fairness: Embed fairness into lot of decision making. Matching on matrimonial site for example. If the data set was trained on fair people, can it stop showing bias against darker people? How do you detect biases in algorithms? How do you take care of this? How do we judge the fairness? What if a woman is searching for a red gown and the results show red gown for a pregnant woman?

3) Data Privacy: GDPR regulations in Europe makes companies accountable. One can’t just scroll or walk around and take data.

Shantanu: As an entrepreneur, I have a lot to say. I ask you to read a lot of books. I touch upon some books you read and you should see on youtube. Why is it that AI consciousness has exploded last few years? Why is human learning not working for us to get the machines to work for us.

Rise of the robots by Martin Fold:  Human beings are going to face stiff competition from AI for the coming decades. We may reassure ourselves that this is nothing new. Look at the youtube “humans need not apply” video. What are these examples. The idea of a transportation industry seems to be fairly routing thing for people who are not in the transportation industry. For example 70 million people work in transportation. They may be replaced. This is a big fear for them. For the first time in history, outsourcing the physical hard labor shifted to mental hard labor. 10 years ago, we paid lots of money to become a radiologist with no contact with patients. Sweet job. Not now. Radiologist as a high quality high earning job is getting disrupted.

Quil is journalism software. 33% of the content in journalism in the world is done by Quil. It can go through the financial statements and write a report. It can take a sport and write like a sport journalist. We kind of accept the competition. One level we are consumers and we love automation. With this hat, we love automation. As job seekers, we hate AI. One of friends in Mumbai ordered dinner. She called Mainland CHina. Then she cancelled it. Then she struggled with the human interface of the person on the phone. Then she ordered on Swiggy. We love as consumers automation on the other end. majority consumers prefer automation. Dichotomy is built directly into us.

AI is great for collaboration. Look at Da Vinchi. I talked to my co-passenger who happened to be a 777 pilot. When I asked him he said, on an 8 hour long haul flight, he said 98% of the times aerolines are on autopilot esp. middle eastern airlines. American and western airlines are Ok upto 90%. Use robot to detonate a bomb.

Companionship: Next big epidemic is loneliness. Lots of psychiatrists get better pay. Elli.q bot is designed to sit on your window sill or desk and be your companion. Firing up skype, downloading things, take you for a walk and all of that. Like Alexa targeted for elderly people.

David Levy, a famous international chess master, mentioned people marry robots in future.

Competition to collaboration to Companionship. Now Convergence. How about enhancing the performance of your brain through say an implant that brings old memories back? Victim of disease. Homedics is a boo you shall read.  Death and disease is a defect that we need to fix.

Rembrand project: He lived in Holland four years ago. They have taken about few hundreds of his paintings esp. portraits of his friends and himself. He got a particular fascination of males with beards and light always fell on the right side of the faces. This group in Holland in collaboration with ING and Microsoft, using 3D paintings, produced new portraits of Rembrandt. It passed turing test too. This is like taking all their work as data and producing new pieces of work. This is not about going back and find experts and look at the work they did and produce their expertise.

Santanu: Humans are the only species in the world who destroy their own habitat. Weather systems modeling and all become easier with AI. Will AI be answer to our pollution. There is plastic island that floats in the ocean. An young person tries to break plastic apart. AI is one revolution and future of destroying plastic is bacteria. Can we create bacteria that eats plastic. This is happening.

PJN: AI can provide waste management tools. We need natural human wisdom. Technology is not there to prevent it.

Karteek: Change has to be you, me and us. AI is just an enabler.

PJN: Democratized super computing is a concept by Jensen Wong. It paid off. Seeing some vision and sticking with it to reap rewards.

Vikas: When students say they can’t do deep learning without GPU, I say constraints are good. When we got 2 GPUs and 100 GPUs, performance never improved. The real ImageNet work is done as two parts as he couldn’t fir it on his gaming laptop.

Santanu: UBI. The future of economic growth is probably jobless growth.Growth of the economy will accelerate not necessarily by employment growth. Consumers go up and producers coming down. Everyone is consuming but who is paying for this. Can we tax these high tech companies and have the government pay to everyone to consume 🙂 There are experiments happening here. India is always an informal economy. Formal employment won’t happen in india easily. Can a government write a check to 95% of the indians? NREGA scheme is that. Government is doing it already. A subsidy is happening at state level too. What does it do to macroeconomics? We don’t know. Only time can tell. We may ask people with UBI to get master’s degree.

Post-facto rationalization can always be given to indian customs. We don’t know that source of origin of such custom or practice.

Q) Is there any hard limit of AI? How far we can go? Can AI eventually pass the real turing test? Moore’s law is already approaching to the end. We only have to depend on the software part?PJN: Compute power is never a limitation. Better algorithms. Today, we can get the turing test pass. The real turing test for machine getting that multi-dimensional awareness of human is 50-100 years away.

Santanu: Turing test is not heavily standardized. Deep Blue beating Gary Kasparov means turing test in the domain of chess already happened. The outer limit to AI is not there same as there is no outer limit to human limit.

Q) There was a news last year that Facebook shut down two chat bots because they started to develop their own language. Why do you say we can’t have AI develop common sense? Someone is predicted to be Gorilla. Face recognition system in China is so scary that criminals on their own confess. Is this a model problem or human problem?Karteek: Human is the problem.

Vikas: Make algorithms more accountable. In the court of law, we can’t say the culprit is neural network black box. Accountability is the key. Humans are good at creative. Only a human can be next Da Vinchi. We are funny and sometimes irrational. That makes us special.