New York and San Francisco, two very different cities, and three very difference AI & ML Conferences. This trip was motivated by the simple fact that many of the visitors to the Informed.AI media websites originate from the USA. So to visit and see first hand what the current topics of discussion are, the latest research, and most interesting applications, was only going to help inform me what the focus should be to continue the expansion and additions to the sites going forward.

New York Academy of Sciences – The 11th Machine Learning Symposium

This was a one day event, hosted by the NYAS, and located in 7 world trade centre, on the 22nd floor.

Three rooms full of poster displays of papers and research, some of which was unpublished so essentially hot off the press as they say. We were asked to not share online the details of the works, due to the fact that some was still to be officially published. This request will make it difficult for me to drill into any specifics here, but I will be able to give an overview of the topics covered without risking any problems for the authors.

Here was the agenda for the day;

The formal part of the event, included three keynote talks, details below, mixed with a series of 5min overviews of some of the key papers and posters that were submitted.

The event was sponsored by American Express, who have a data science group and were actively recruiting at the event.

The three keynote talks where;

  1. Challenges in Privacy-Preserving Data Analysis by Kamalika Chaudhuri
  2. Machine Learning for Individualised Healthcare by Suchi Saria
  3. Applications of Learning Theory in Algorithmic Game Theory by Tim Roughgarden

And here is a list of the poster abstracts

In terms of review, and starting with the three keynote talks;

The first talk highlighted the privacy triangle which shows the tradeoffs of data size and accuracy. The topic of Differential Privacy was discussed and the method of Markov Blankets highlighted. Also a paper was referenced called Pufferfish Privacy. A process of Model-Infer-Criticise was also discussed. Unfortunately the methods only work certain models. Overall the    summary about privacy was that it is difficult to provide complete and total privacy, I liked the quote given from George E P Box (1919-2013) “All models are wrong, some are useful”

The second keynote talk on Individual Healthcare highlighted the challenges of data collection, and in particular the process of selecting appropriate features.

The third talk, on learning theory reflected on the price of anarchy and optimal auction design and of course mentioned the Nash equilibrium. It also showed that the Ebay auction design is as close to the perfect auction design as you get.

In terms of the posters, as you will see from the abstracts, there was a wide range of topics covered. but as expected alot of focus on NLP methods, feature extraction automation, and reinforcement learning methods.

The event was closed with a drinks reception which allowed for additional networking and discussions. Overall a really good event, with a good mix of academic and commercial research being showcased.

AI by the Bay

AI by the Bay was a three day conference, each day having a different focus. Hosted in the bay area of San Francisco at a venue called the Pulse, which was a great open space, including a mezzanine and root terrace, which hosted the evening parties and lunchtime activities.

Alex, the founder of AI by the Bay, gives the opening introduction to the conference.

The Three Day Agenda is shown below:

Essentially the format of the event was 30-40min talks, with a panel event at the end of the day. The third day was slightly different as we have three panel events one before lunch, and two at the very end of the day.

The event was very well organised, and hosted by Alexy, who runs a number of other conferences and meetups in the Bay area.

Day 1 – Oh Hai AI

The first day, for me was the most interesting, as it covered more general AI and ML topics. Overall the day had alot of focus on what it takes to build a fully working ML system, rather than just building a ML model that just shows some interesting data science analytics.

A number of different platform designs where presented:

One of the best talks in the morning session was by SalesForce, on Attention and Memory. Which started with RNNs for language translation, but then discussed Episodic Memory and the QA methods using images.

The afternoon highlights was a demonstration of BachBot which has now been integrated into the Google Magenta project.

Using NNs as generative models to essentially generate content, and more specifically artistic content is very interesting area, and one which I think we will see more and more of in the coming months and years.

Talks by Domino Data Labs and IBM closed out the talks for the first day.

The Event Sponsors are listed below:

The first day panel event was about taking AI from Idea to Customer;

As a footnote for the first day, I was live tweeting throughout the day, so you can also look at my tweet history to see more on the days at @DrAndyPardoe and you can also see the twitter network map to see interactions during the day;

Day 2 – Self Driving Cars

The second day was probably the least interesting for me, however, it was good to see how much energy there was for this topic. Talks from NDIVIA, Otto, Comma, PolySync, Velodyne and others. Two key discussions during the day, one on the use of Lidar, and the second on Self-Driving Trucks.

The topic of Lidar (Laser based Radar), was interesting as all of the companies except one (Telsa) are using Lidar. I have to say, based on what I heard, I understand why Elon Musk has decided not to use Lidar. Reasons seem to include, Expensive, limited manufacturing capacity, and while it give good 3D high resolution images, it has limited range and so is limited in terms of the speed of the vehicle when it can be used.

The second topic, saw three companies include Otto (recently bought by Uber) that are focused on Lorries (Trucks). The last company, who were a very early startup, had a very interesting angle on this application, and I think we will be seeing more of them in the future. They essentially want to try to completely remove the driver from the trucks and replace them with a call-centre remote control setup.

Day 3 – The Future of AI

The third day was opened by keynote speaker Stuart Russell, professor at UC Berkley, who Andrew Ng studied under and who is running the Centre for Human-Compatible AI. Prof Russell has some interesting views on the future of AI

Stuart has a view of concern about controlling the robots to ensure they don’t miss interpret what we truly want and deliver this by ensuring there is a level of uncertainty programmed into the robots. A very thought provoking talk, and set us up for a very intensive day of discussions.

The next talk was on Advances in Deep Learning for Mathematical Theorem Proving, which focused on abstraction and reasoning. Essentially moving away from simple pattern matching to more ability with encoding abstractions to allow better reasoning.

Next was a talk from Bradford Cross founding partner of DCVC on the topic of Venture Capital investment in AI startups. The talk used some data from CBInsights to highlight potential areas of both saturation and target opportunities. The focus on applying AI to vertical markets. This was followed by a panel discussion.

The afternoon was opened with a talk on AI Vision and a training course by Then one of the engineers from OpenAI demonstrated the use of the Gym Environment for training agents. We then has some focus on Chatbots with another panel discussion.

The last panel event which essentially closed the conference covered the Future of AI, with both Peter Norvig and Stuart Russell. The panel were asked what keeps them awake at night, and what do they want to see happening with AI.

We also heard that a 4th edition of the AI book by Norvig and Russell will be published soon, to include a 19th chapter on Deep Learning.

The final roof party was hosted and augmented by a number of startup companies from the portfolio of the party sponsor, DCVC.

Machine Learning at Community Banks

This was a one day event, organised by Startup.ML and hosted at the Bloomberg building.

The agenda was as follows;

  1. Collateralized Lending using Machine Learning, Karén Chaltikian, Lending.AI
  2. Machine Learning for Delinquency and Default Risk, Dhruv Madeka, Bloomberg
  3. KYC and Fraud Prevention in Crypto, Soups Ranjan, Coinbase
  4. Tips and Tricks for Machine Learning Projects, Brendan Herger, Capital One
  5. Self-Supervised Machine Learning, Arshak Navruzyan, Startup.ML
  6. FinTech Panel, moderated by Clifford Tong

The first talk on collateralized lending, discussed in detail the various sources of data that could be used, especially during the lifetime of any lending, but highlighted the difficulties of using such information.

The second talk from Bloomberg, took a challenge from kaggle and showed the different approaches and models to get improved results, including the use of blending models. This demonstrated the complexities of achieving high accuracy modelling with a real world dataset.

The talk from Coinbase, discussed using ML to detect fraud, and how they dynamically adjust credit limit of customers based on real-time information they source.

The tips for ML Projects talk given by Capital One, had an interesting take on running POCs.

The final talk gave a high level summary of self-supervising machine learning which is a new concept developed by Startup,ML, details on this were limited, but more will be made public very soon.

The panel discussion on FinTech gave an insight into what FinTech’s are doing with Community Banks and how they view them compared with the larger national and international banks. It seems FinTechs are able to quicker wins with the Community banks allowing the FinTechs to continue to develop their offering which in parallel they work with larger banks, but with the expectation that engagement with the national banks will take alot longer to materialise.


This concludes the summary review of the three conferences, and I know in places I have only scratched the surface on the details that were shared during the talks and discussions, but hopefully this has given a good summary of the events, which should be useful for people that didn’t have the opportunity to attend.

For the AI by the Bay conference I should be able to update this blog later with a link to the full set of slides and videos of the talks, which if you want more information on a specific talk will of course be the best source … watch this space.

I have had the opportunity to see some fantastic new work performed by both students and researchers, and to meet so many amazing people working in this industry. I very much enjoyed by time in America and look forward to visiting again soon.

If you would like to learn more, why not attend my next meet up or just follow me on twitter