From building a search engine for lectures, I've found that there is more interest in artificial intelligence, as it the discipline migrates from research applications to highly visible commercial products. It can be difficult to find good introductory material, especially for people who are interested but non-technical.
Much of the research is very math-heavy, and productizing it can require software engineering and operational expertise. Further, the discipline is fragmenting around use cases and data formats - for instance, there are products that focus on working with images, sounds, or text, and many AI-based products are actually intended to be building blocks, rather than complete solutions.
In this post, I've collected several talks I've found that have helped me understand the connections between business and technical problems.
AI: A Return to Meaning - David Ferrucci - 65 minutes
I saw this talk by David Ferucci at Philly ETE. He led the team that created IBM's Jeopardy-winning Watson software. In this talk he covers how computer scientists have teach computers to grapple with natural language problems. Early attempts tended to be rule oriented (e.g. building databases of grammatical structures), but with time have moved to probability models. One of the compelling parts of this talk is a discussion of how Watson deals with ambiguity in Jeopardy games.
Anthony Levandowski: Google Driverless Cars
This talk covers some history and motivations for the development of driverless cars, including experimentation into driverless motorcycles and tractors. This started with early DARPA research, with a more recent focus on urban driving. This will give you a window into some of the non-technical problems around AI, as the investment into engineering and legal issues around autonomous vehicles is immense and complex.
This is an older keynote, also from Philly ETE. The speaker works in online advertising - she talks about how internet advertising evolved to a market-based system, and how computer scientists think about buying the best ad space and combatting fraud. I thought this was also a helpful perspective of what researchers do on a day-to-day basis.
One of the challenging of machine learning is understanding what the computer is actually "learning." This talk focuses on a unique angle - novel visualization techniques to see inside the machine's mind. Some of this assumes a little bit of technical background, but if you stick through, they cover some interesting edge cases for where AI fails.
Amazon Machine Learning: Empowering Developers to Build Smart Applications
This is unfortunately a bit of a sales pitch for an Amazon product. It demonstrates what Amazon feels is useful for people who want to incorporate machine learning into their applications. This seems to be aimed at people who use spreadsheets to make business decisions (e.g. for an e-commerce site) but have a data volume beyond what Excel can handle. Since this is a product demo, it will give you an idea of some practical problems you'll face, like making sure your results are valid and repeatable, and controlling the balance of false negatives/positives.
More Resources
If you're interested in exploring these topics further, I've also found the AlchemyAPI demo site helpful for text analysis, and the PyImageSearch blog for learning about vision / image recognition problems.
Have any good talks you can recommend? Please comment below, and I'll include them in https://www.findlectures.com.
Top comments (2)
Looks like some great resources. Thank you. One thing I'm struggling with since I've started paying attention to all the ML, deep learning stuff is that ML and AI are used interchangeably.
For every problem, it seems like, the first reaction is to wonder how to use ML to solve it.
It may be the cynic in me but I feel like few years ago every conversation or solution to a problem would lead very quickly to 'Big data'
I think where I'm getting at is (finally!) What resources are there to help me clearly understand the scope of ML and AI and how to develop my thinking to objectively identify problems where they provide a good solution
Side note: there was recently a post which made it to HN front page about how deep learning isn't the best solution for a small dataset and a counter post later about how it is..
Agreed. There is definitely a hype cycle on this (I have a school textbook for "data mining", which was also a precursor) What I like about these talks is they are primarily case studies of projects, so you can see how the team broke down a problem, as opposed to focusing on specific, tactical algorithms or techniques.