In September 2016, the ReWork team organised another deep learning conference in London. This is the third of their conferences I have attended and each time they continue to be a fantastic cross section of academia, enterprise research and start-ups. As usual, I took a large amount of notes on both days and I’ll be putting these up as separate posts, this one covers the morning of day 1. For reference, the notes from previous events can be found here: Boston 2015, Boston 2016.
We’re all starting to get a bit blasé about self driving cars now. They were a novelty when they first came out, but even if the vast majority of us have never seen one, let alone been in one, we know they’re there and they work1 and that they are getting better with each iteration (which is phenomenally fast). But after watching the formula 1 racing, it’s a big step from a 30mph trundle around a city to over 200mph racing around a track with other cars. Or is it? Continue reading Formula AI – driverless racing
As part of a few hours catching up on machine learning conference videos, I found this great talk on what can go wrong with machine recommendations and how testing can be improved. Evan gives some great examples of where the algorithms can give unintended outputs. In some cases this is an emergent property of correlations in the data, and in others, it’s down to missing examples in the test set. While the talk is entertaining and shows some very important examples, it made me realise something that has been missing. The machine learning community, having grown out of academia, does not have the same rigour of developmental processes as standard software development.
Regardless of the type of machine learning employed, testing and quantification of results is all down to the test set that is used. Accuracy against the test set, simpler architectures, fewer training examples are the focal points. If the test set is lacking, this is not uncovered as part of the process. Models that show high precision and recall can often fail when “real” data is passed through them. Sometimes in spectacular ways as outlined in Evan’s talk: adjusting pricing for Mac users, Amazon recommending inappropriate products or Facebook’s misclassification of people. These problems are either solved with manual hacks after the algorithms have run or by adding specific issues to the test set. While there are businesses that take the same approach with their software, they are thankfully few and far between and most companies now use some form of continuous integration, automated testing and then rigorous manual testing.
The only part of this process that will truly solve the problem is the addition of rigorous manual testing by professional testers. Good testers are very hard to find, in some respect harder than it is to find good developers. Testing is often seen as a second class profession to development and I feel this is really unfair. Good testers can understand the application they are testing on multiple levels, create the automated functional tests and make sure that everything you expect to work, works. But they also know how to stress an application – push it well beyond what it was designed to do, just to see whether these cases will be handled. What assumptions were made that can be challenged. A really good tester will see deficiencies in test sets and think “what happens if…”, they’ll sneak the bizarre examples in for the challenge.
One of the most difficult things about being a tester in the machine learning space is that in order to understand all the ways in which things can go wrong, you do need some appreciation of how the underlying system works, rather than a complete black box. Knowing that most vision networks look for edges would prompt a good tester to throw in random patterns, from animal prints to static noise. A good tester would look of examples not covered by the test set and make sure the negatives far outweigh the samples the developers used to create the model.
So where are all the specialist testers for machine learning? I think the industry really needs them before we have (any more) decision engines in our lives that have hidden issues waiting to emerge…
Following my post on AI for understanding ambiguity, I got into a discussion with a friend covering the limitations of AI if we only try to emulate ourselves. His premise was that we know so little about how our brains actually enable us to have our rich independent thoughts that if we constrain AI to what we observe, an ability to converse in our native language and perform tasks that we can do with higher precision, then we will completely limit their potential. I had a similar conversation in the summer of 2015 while at the start-up company I joined1– we spent a whole day2 discussing whether in 100 years’ time the only jobs that humans would do would be to code the robots. While the technological revolution is changing how we live and work, and yes it will remove some jobs and create others just like the industrial revolution did and ongoing machine automation has been doing, there will always be a rich variety of new roles that require our unique skills and imagination, our ability to adapt and look beyond what we know. Continue reading Evolving Machines
I’ve been with my current company for 9 months as Chief Information Officer and had responsibility for everything technical from production systems down to ensuring the phone systems worked and everything in between. The only technical responsibilities not under me was the actual development and QA of our products. CIO is a thankless role – when everything is going right, questions are raised over the size of the team and the need to replace servers and budget for new projects. When something breaks, for whatever reason, you are the focus of the negativity until it is resolved. The past 9 months have been a rollercoaster of business needs, including many sleepless nights. However, I can look back on this knowing that when I do finally get around to writing about my experiences as a woman in IT I will have a lot of fun stories for the CIO chapter1. While I didn’t have the opportunity to finish off as many of the improvement projects as I would have liked, I’ve built up a fantastic team and know that they’ll continue to do a fantastic job going forward.
Last year I wrote a post on whether machines could ever think1. Recently, in addition to all the general chatbot competitions, there has been a new type of test for deeper contextual understanding rather than the dumb and obvious meanings of words. English2 has a rich variety of meanings of words with the primary as the most common and then secondary and tertiary meanings further down in the dictionary. It’s probably been a while since you last sat down and read a dictionary, or even used an online one other than to find a synonym, antonym or check your spelling3 but as humans we rely mostly on our vocabulary and context that we’ve picked up from education and experience.
The following tweet appeared on my timeline today:
The Turing test is like saying planes don’t fly unless they can fool birds into thinking they’re birds. (h/t Peter Norvig) #AI
— Pedro Domingos (@pmddomingos) July 19, 2015
Initially I thought “heh, fair point – we are defining that the only true intelligence is described by the properties humans exhibit”, and my in-built twitter filter1 ignored the inaccuracies of the analogy. I clicked on the tweet as I wanted to see what the responses were and whether there was a better metaphor that I could talk about. There wasn’t – the responses were mainly variants on the deficiencies of the analogy and equally problematic in their own right. While this didn’t descend into anything abusive2, I do feel that the essence of what was trying to be conveyed was lost and this will be a continual problem with twitter. One of the better responses3 did point out that cherry-picking a single feature was not the same as the Turing Test. However, this did get me thinking based on my initial interpretation of the tweet.
In order to answer a big question we are simplifying it in one way. Turing simplified “Can machines think?” to “can machines fool humans into thinking they are human?”. Continue reading Can machines think?
There’s a lot of money, time and brain power going in to various machine learning techniques to take the aggravation out of manually tagging images so that they appear in searches and can be categorised effectively. However, we are strangely fault-intolerant of machines when they get it wrong – too many “unknowns” and we’re less likely to use the services but a couple of bad predictions and we’re aghast about how bad the solution is.
With a lot of the big players coming out with image categorisers, there is the question as whether it’s really worth anyone building their own when you can pay a nominal fee to use the API of an existing system. The only way to really know is to see how well these systems do “in the wild” – sure they have high precision and recall on the test sets, but when an actual user uploads and image and frowns at the result, something isn’t right. Continue reading AI for image recognition – still a way to go
So, day one of the ReWork Deep Learning Summit Boston 2015 is over. A lot of interesting talks and demonstrations all round. All talks were recorded so I will update this post as they become available with the links to wherever the recordings are posted – I know I’ll be rewatching them.
Following a brief introduction the day kicked off with a presentation from Christian Szegedy of Google looking at the deep learning they had set up to analyse YouTube videos. They’d taken the traditional networks used in Google and made them smaller, discovering that an architecture with several layers of small networks was more computationally efficient that larger ones, with a 5 level (inception-5) most efficient. Several papers were referenced, which I’ll need to look up later, but the results looked interesting.