Monday, January 25, 2016

Interview with Antonio Roldao, CTO/Co-Founder of

Who are you?

I am the co-founder and CTO at Muse.Ai. Before starting this company, I was part of a technology consultancy team within Morgan Stanley where I pioneered a hardware-based ultra-low-latency trading system, and implemented a Big Data System for network analysis, data visualisation, and insight generation from large amounts of usage data. Prior to Morgan Stanley, I worked at JP Morgan in an elite team developing a state-of-the-art risk management and trade-life-cycle platform (Athena). I hold a Ph.D. in accelerating Scientific Computations using Iterative Methods on dynamically reprogrammable hardware (FPGAs) from Imperial College London. In addition to having both hardware and software operating in space onboard the Giove-A satellite, I have also developed WebCanvas - The World’s Largest Collaborative Online Painting.

What does do?

Muse.Ai is a video platform that allows users to view specific segments of video based on keywords. This is achieved by bringing together a number of leading-edge algorithms and methodologies to extract and index information from videos allowing for fast recall and replay of specific segments of interest. For example, if you wanted to recall a particular visual and auditory memory such as “when JFK talked about the going to the moon” or “when Steve Jobs talked about mind bicycles” the system will output all video sections that match these criteria. These videos are sorted by relevancy and start playing immediately in the relevant part (e.g. VideoRank).

What caused you to start

The concept for Muse.Ai came out of a personal frustration when I had to go through hours of recorded lectures to find a specific moment when the instructor showed the derivation of a certain formula. If I had already developed the technology of Muse.Ai then, I would have saved myself hours of skipping through video to find that one specific section that I was looking for.

What makes an AI startup?

We use Artificial Intelligence at many different levels of our technology stack, for example: in order to effectively index and retrieve particular sections of videos, the system needs to first ingest the video. This ingestion decomposes each video into its constituent parts (sound, image, motion) and performs deep analysis. This analysis involves a multitude of AI systems that do a myriad of tasks including: objects and individuals recognition, words and symbols recognition, activities recognition, speech to text, etc. The output of all these analyses is then indexed using a custom-built engine for ultra-fast look-ups. The results of these look-ups are once again sorted using another adaptive AI system, akin to a recommendation system that optimises the output for the user’s context/interests.

Why is there so much excitement about AI startups?

I think the excitement comes from the multitude of factors that are enabling new AI start-up. Some examples of enabling factors are access to unprecedented amounts of computing power (e.g. Cloud, GPUs, FPGAs, etc.), access to massive amounts of data (e.g. Wikipedia, YouTube,, etc.), readily available algorithms, libraries, and APIs (e.g. on GitHub, Cloud Services, etc.), miniaturised high-precision sensors (e.g. MEMs) and wireless high-bandwidth connections (e.g. 4G in Smart Devices). There is also a lot of innovation in terms of robotics (particularly with 3D printers, drones, etc.) that are opening-up new and exiting opportunities.

In my view, these new opportunities can be divided into two categories: those that replace existing services (e.g. self-driving cars, translation tools, etc.), and those that create services, products, and jobs, that never existed before, and weren’t even possible. In the same way that airplane pilots, TV-presenters, and programmers did not exist 100 years ago, just think what new jobs may be created next! As a passionate technologist, I am particularly excited and curious about this second category.

Describe's AI.

Our AI can be described as a set of small and very weak AIs that come together to build a stronger AI. The smaller AIs have algorithms that do relatively trivial tasks such as scene change detection all the way-up to recognising what activity is being performed on a set of images, and by whom.

Our architecture is very much inspired by biology where each type of cell is dedicated to a specific task, but combined create first organs and then a functional organism. For example, we could metaphorically say that our system has “organs” that ingest videos, break them down into their constituent parts and distribute them to the appropriate processing subsystems. The processing subsystems then produce refined information that is stored in another critical subsystem. There is of course also a sub-system that cleans up any unwanted residues.

How do you measure and communicate the quality of your AI?

In our particular case it is relatively easy to judge if the AI produced a correct answer because there is a visual and auditory output. This means that we can rapidly inform the system that the answer was not as expected and iteratively improve both the recommendation and the ingestion systems.

Having said that, one of the findings that really fascinates me is when the machine is actually “right” or more sensible than Humans. For example, English is supposed to be a phonetic language (i.e. we write and read the sounds we make) but in reality there are many edge cases where the sounds we make and symbols we use to represent those sounds are completely illogical (e.g. in the case of Heterophones, and Homophones). When the AI does not know about these exceptions and edge cases, it comes up with logical but wrong answers!

Can you share something awesome that your AI has done or that you have been able to do with your AI that would surprise most people?

There have been many surprising findings, but one that was particularly striking was what happens when we get a large array of short videos on a topic. Having this ability to SkipSearch in-between videos directly to where some concept is relevant allows one to identify patterns and understand certain concepts in many different contexts in a rapid-fire way that was not possible before. For example, looking up the word “love” one is able to see and hear a myriad of contexts in which this word appears, and in this particular case the most striking insight was its association with the concept of forgiveness. When I came to think about it, it really makes a lot of sense - we are willing to forgive much more from those that we love.

When you talk to investors, do they care about your AI or do they care that you are solving a business problem? How about customers? How about the media?

Investors that have seen our prototype have become very excited about the range of applications of our technology. They are particularly interested in how it can be applied to help professionals perfect their understanding and techniques (e.g. how to perform a particular medical procedure), or to allow for custom tailored entertainment (e.g. show me goals by Ronaldo). This wide range applications means that they are confident that we can target many business problems.

In terms of customers, my co-founder and I are the first enthusiastic consumers of our own technology. We love it to the point that we get annoyed when these capabilities are not present on other platforms.

In relation to the media, we have not yet launched our service and for the time being would rather keep a low profile while we perfect our systems.

Where do you go to get information about AI startups?

There are many great sources of information online, and offline. On-line, I tend to follow a number of twitter accounts, blogs, reddits, and AngelList. Off-line, I tend to learn about exciting start-ups through publications like Wired, AI-related meet-ups, and specialised conferences (PyCon and NIPS, for instance) and meeting with leading academics in the field (such as Andrew Ng, Yann LeCun, Geoffrey Hinton).

Make a prediction about the future of AI startups.

As I mentioned previously, in my view, there are two types of AI-startups: those that automate some sort of existing process, and those that create completely new services. The first category will naturally generate a lot of resentment, in particular from those who will have to adapt into new jobs. The other type of AI start-up will open-up new opportunities that we cannot yet imagine.

I also think that there will be an interesting shift from companies focused on reactive AI systems (ask/reply-type) to pro-active AI systems (automatically inform/suggest at opportune moments). These pro-active systems can potentially massively improve quality of life.

No comments:

Post a Comment