Bhat's Notes

Moderator: Sainath Gupta

Sundara R Nagalingam, Head of Deep Learning, NVIDIA   

Mahesh Yellai, Founder & CEO, Infruid Labs  

Visesh Handa, Head of Service Delivery, Arthayantra

Visesh: Arthayantra is into personal finance (insurance and investments).

Mahesh: We are in enterprise space. We build AI wizard (Q&A bot for business insights). It makes  non-technical users get answers about business insights.

Mahesh) AI is ensuring machines think and act at the same level as humans. AI is ensuring machines mimic humans.

Sundar) One methodology of AI is deep learning. Create models and train models. I showed my 2 year old son a red apple and told him it is called apple. When a green apple of different size is shown to him, he also identified it as apple. He did this because he got trained first and then learnt from that information and identified green apple. This is what deep learning does. We have to give large number of training data instances so deep learning model can train on and make appropriate decisions. We only train the computer but don’t tell the computer. Deep Learning is a subset of AI.

Visesh) Can you see an individual behaviours and continuously learn their risk taking patterns? AI is about continuous learning from human behavior. Ability to learn through patterns.

Mahesh) Google’s captcha. Earlier we got to type a number to find a human in the picture. Google captcha can tell automatically whether you are a human or not. Just identify the image and check the box so google AI knows you are human.

Sundar)

1) Image/Video Analytics

2) Speech

3) NLP

Don’t key in demographics. Just load the picture and AI gets demographics such as gender and age.

When someone walks into Target store, where do his eyeballs look? Which ones did he touch and looked? How many items got into his basket? This is all AI.

Video surveillance: Out of 10000 people between 9am and 10 pm, only two people looked suspicious. This info got to be picked and police got to take action. This is AI based security surveillance. Big problem to solve.

Healthcare: Sitting all day in front of a monochrome monitor at an MRI lab gives radiologist fatigue. Machine can see aberrations in the scan. Algorithm refers to the knowledge base and identifies the area of the scan that radiologist just got to approve.

In ATM chamber, incidence of interest such as pulling out of a gun got to be flagged.

Visesh) Fraud detection. How do we do it at a bank? All pattern recognition. Fraud detection in financial sector is incrementally relevant. Another area is loan under writing. No regular salary and credible history are reasons why banks decline loan yet the have lots of NPAs. How regular banks have higher NPA than Mudra Bank NPAs that are not based on any collaterals. Can we identify a pattern from these Mudra banks?

Mahesh) Speech to text to actions! Voice is faster and natural for humans. We see lot of inroads by STT (Speech to Text) and STA (Speech to Action) in voice. In the enterprise world we don’t see that happen though we could see it through Siri and all. We want to start from text and move to speech to text and action. AI impacted more of consumer space than enterprise space.

Even google tried to do voice search but couldn’t do it well. Only healthcare made some inroads.

Q) Growing power of graphics card?

Sundaram) NVidia’s GPUs are at central stage or AI. This development is due to NVidia presence and research. More data needs more security. Deep learning prediction now is higher than average human predictability. NVidia is a de facto stardard now.

Visesh)

1) Average human makes 26000 financial decisions. How do we optimise these decision so we can fare well? We use decision trees and neural network and zillions of computations to arrive at this.

2) Predictive modeling: At 99% accuracy we predict how individual makes choices based on some criteria.

3) Fraud detection and risk management: Identify individual risk behaviours. When we study their behavior, we learn more insights through continuous improvement.

Mahesh) initial goal of AI is to mimic humans. How humans look at symbols and draw meanings. In the industry, we see the rote methods take place. Neural networks are still rote method. They don’t look at symbols to draw the meaning. This is the strange thing. Image processing and video analytics are not chasing human behavior. These algorithms don’t mimic human thinking. Can we translate human behavior through machines? Humans interface with machines and make thought in mind manifest the product by itself in front. Can you make interface of the human to the machine more intelligent where humans do less and machines to more. If Amazon go takes over, you don’t have retail jobs.

Interfaces to machines is a broad opportunity for career. Job displacement and new jobs are bound to happen.

Sundar) Healthcare is big. It is associated with human life and most precious thing. Identification of cancel cells, diabetic retinopathy, scans and all are going to have a big impact. This is not matured as of now. A big growth area.

Ecommerce is kind of matured. One novel idea is recipe book app. App takes pic of refrigerator that scans the vegetables and recommends possible recipes you can have. Be imaginative to come out with interesting products.

Finance industry is another big industry. Say signature verification at bank. Validity of the date on check. Checking the amount on the check and words. This is getting fast automated.

Today I type in ICICI online where is nearby ATM. AI can do consumer targeted advertisement in chat bots. Lots of work is happening here. Sentiment analysis and buy sell decisions through structured data. Nice technology called graph analytics where sentiment analysis is used. When many unrelated factors influence a decision such as wife likes the shirt, conventional Bayesian model won’t work and here graph analytics could help.

Bad news about a bank from an article on the other side of the globe could trouble its stock, though bank’s financials look good.

Visesh) If a learning is required at scale and a decision got to be taken, that is opportunity for AI. Turn left or right? Learning required yes. Huge scale? yes. Similarly, buying a shirt? Is there a learning required based on ambiguous set of large scale data, then AI can help make better decisions.

Mahesh) Enterprise got to increase revenues or reduce the cost. Look at ways in which AI can impact top line or bottom line. Look at jobs a person does and find t jobs with high friction and try to reduce that burden through AI. How can you make grocery buying experience friction less?

Sireesh) Industrial revolutions happen like this. This is not the first time. Do humans always take right decisons? No. Same with machines. I am aware in a big corp where AI based spam control mdeol thought spammers are right and non-spammers are wrong. Machines can make mistakes at scale too. Be careful.

Sundar) AI and DL are glorified. Scary. Where would be the real inflection point? Is this godsend? There is so much attraction towards it. I feel some apps are getting crowded.

Technology can never lead to unemployment in the long run.

Mahesh) Displacement of jobs will be there. Mis-classification is a problem. Targets sent coupons for pregnant women to a family and the father of the teen fought with Target manager. Before daughter recognised she was pregnant, Target figured it out.

Sainath Gupta: 33% of the employment in USA is through trucking industry and AI can wipe it out.

Ranjit Q) I am interested in analysing conflicting human perespectives. Your sense on singularity and end game of AI?

Sundar) Many things that touch human life would be done by AI. Final approval authority will still be human being.

Mahesh) Machines are becoming more intelligent and humans are becoming less intelligent. Which party to vote for is fed by the newsfeed in your social media.

We can’t judge on this.

Sainath: Costs come down. IBM Watson with Manipal could bring cancer second opinion price to Rs8000 from Sloan Kettering price of $2000.

Sundar) Top universities are doing big research. Deep learning is not yet offered as course except for few. Second tier schools are aware. Majority of the colleges in India ignore or ignorant. Technology will catch hold of students even if students don’t about them.

Mahesh) Tribal sharing in 90s to democratisation through web in 2000s changed how Indian applicants found US schools and got admissions.

Bangalore based CrowdAnalytics is popular like kaggle. Traditional engineering colleges don’t need you to educate, net and MOOCS can do.

Sundar) Imagenet is library of images. In imagenet challenge of identifying the image, accuracy went from 73% accuracy in 2011 to 95% in 2016. Even human could only get 92%. Machine prediction is better than average human. AI is one of the great technologies that revolutionized computing.

Netflix got very sophisticated movie recommendation algorithm. Crowd sourcing led to someone win a million. Some smart guy figured out anonymous data to de-anonymised data. They stopped awarding anymore. GE also does it.

Q) Any source for training  data?

Data is sustainable competitive edge and why would they like to put in the public world?