Evening Standard: TfL may use AI to catch fare dodgers

[Haywain]

I've only read the newspaper article so maybe the reports explain this,

The article linked to in post #20 is quite good, and gives fairly thorough explanation.

 

[AlterEgo]

James O'Malley is a good writer and I'm pleased to see his article. I think I'm a bit less cynical about AI than I was before; this clearly can be used for incremental gains which the railway so badly needs.

 

[Haywain]

James O'Malley is a good writer and I'm pleased to see his article. I think I'm a bit less cynical about AI than I was before; this clearly can be used for incremental gains which the railway so badly needs.

Yes, it's good to read something written in a well balanced way.

 

[AdamWW]

The article linked to in post #20 is quite good, and gives fairly thorough explanation.

It does, and if you read footnote 6 you'll see that he is also sceptical that a training set consisting of a single person would work.

 

[Meerkat]

A key issue that can also exist is that, if applied in certain ways, it can worsen any biases that may exist; for example, if the data feeding the model is based on enforcement action, and the model is used to inform future enforcement action, it can exacerbate any biases in the original dataset.

Maybe I am cyncial about human nature but I would say that whilst AI can learn biases its probably easier to reprogram them out than it is to do that for a human!

 

[Towers]

Interesting comments regarding gateline/barrier experiences & staff. The elephant in the room here is of course the fact that even if AI detects fare dodgers, be they coming in or going out, you ultimately still need the staff on the ground to act upon that information. And gateline staff will likely be very well aware of much of it already, someone tailgating or even barging through a gateline is hardly difficult to spot, but again they simply cannot act upon it. I suppose you might use the information to plan targeted operations in high risk areas, but beyond that it’s difficult to see what AI can achieve here. The platform stuff however sounds more promising!

 

[jon81uk]

Interesting comments regarding gateline/barrier experiences & staff. The elephant in the room here is of course the fact that even if AI detects fare dodgers, be they coming in or going out, you ultimately still need the staff on the ground to act upon that information. And gateline staff will likely be very well aware of much of it already, someone tailgating or even barging through a gateline is hardly difficult to spot, but again they simply cannot act upon it. I suppose you might use the information to plan targeted operations in high risk areas, but beyond that it’s difficult to see what AI can achieve here. The platform stuff however sounds more promising!

I would expect anything found by AI would go the the revenue inspectors, not the gateline staff.

 

[Towers]

I would expect anything found by AI would go the the revenue inspectors, not the gateline staff.

Intelligence may do yes, but that’s of little benefit in the immediate short term of course.

 

[Haywain]

Interesting comments regarding gateline/barrier experiences & staff. The elephant in the room here is of course the fact that even if AI detects fare dodgers, be they coming in or going out, you ultimately still need the staff on the ground to act upon that information. And gateline staff will likely be very well aware of much of it already, someone tailgating or even barging through a gateline is hardly difficult to spot, but again they simply cannot act upon it. I suppose you might use the information to plan targeted operations in high risk areas, but beyond that it’s difficult to see what AI can achieve here. The platform stuff however sounds more promising!

If AI can identify aspects of behaviour prior to getting to the gates then that will help staff to take appropriate action at that point. And if it gets rolled out more widely it could be recognising those people on the way in to the system and flagging them before they get to leave.

 

[Towers]

If AI can identify aspects of behaviour prior to getting to the gates then that will help staff to take appropriate action at that point. And if it gets rolled out more widely it could be recognising those people on the way in to the system and flagging them before they get to leave.

It could, but again you require appropriate staff on the ground to intervene every time for that to be effective. As others have already suggested, that is unlikely to fall under the remit of gateline colleagues, so LUL would need a massive increase in revenue protection or other appropriate staff if there was an intention to make widespread use of AI for this purpose.

 

[fandroid]

Whilst musing about the feasibility of re-introducing bendy buses to London, I did ponder about cameras linked to to the touch pads at all the doors, to identify the free-riders and alert drivers and revenue teams.

The way AI was used in the trial at Willesden Green to spot gate-crashers brought these thoughts to mind again, and shows how persistent offenders could be identified and potentially ambushed

 

[enginedin]

its probably easier to reprogram them out

It depends.

AI is (still) a buzzword, and means a lot of different things to a lot of different people. In many cases, when people say AI, they actually mean ML (machine learning). My (professional) opinion is that this trial is ML, not AI. ML models are just(!) complex probabilistic models. In the most simple case of detecting pictures of dogs and cats, the model needs to be trained with labelled images (typically thousands, or tens of thousands), and then given an unlabelled image, it will give a probability score for the image being a cat or a dog.

Given the output is a probability of a given outcome, it is actually quite difficult to program biases out of ML models. It would (probably, depending on the model) require retraining of the entire model with updated training (and validation) data.

 

[AdamWW]

It depends.

AI is (still) a buzzword, and means a lot of different things to a lot of different people. In many cases, when people say AI, they actually mean ML (machine learning). My (professional) opinion is that this trial is ML, not AI. ML models are just(!) complex probabilistic models. In the most simple case of detecting pictures of dogs and cats, the model needs to be trained with labelled images (typically thousands, or tens of thousands), and then given an unlabelled image, it will give a probability score for the image being a cat or a dog.

Quite.

So it seems somewhat implausible that they taught it to spot weapons based on images of a single person.

It seems to me that the term AI is now in general use just to mean machine learning.

Presumably if at some point someone can manage something that is actually intelligent rather than doing a very impressive job at appearing to be intelligent, we'll need a new name for it.

 

[Bletchleyite]

All current AI is probabilistic machine learning. Including ChatGPT which works with a model along the lines of "probabilistically based on the learning corpus what should the next word be".

 

[AdamWW]

All current AI is probabilistic machine learning. Including ChatGPT which works with a model along the lines of "probabilistically based on the learning corpus what should the next word be".

Quite.

Shouldn't we save the term AI for something a bit more than that?

Although it's amazing how well something like ChatGPT does simulate intelligence given how it works. I wasn't so impressed when it spent a while explaining to me how to do wheelies on a unicycle though.