AI

Lets see if I can explain one of the most scary areas of AI  — neural networks and why they are scary. I’ll probably fail, but I have warned you!

Imagine if you will a number of quite complex wiring blocks, but each has the same rough layouts with the inputs going in on A and the outputs on B. Lets imagine this is a sensor net in which the sensor inputs go from no input up to 100% with all the possible points in between. Lets also have the outputs run from undetected up to definitely detected with all the values in between. Easy so far so lets make the relationship non linear. In a military context this could be a PIR detector for an autonomous gun firing sentry.

Now we need to decide whether that was a leaf, rabbit or person, so feed in the output from the first layer into a second layer that detects the size of the object depending how many PIRs were simultaneously triggered and throw out (set to zero) all the little objects. Still easy and still deterministic.

Lets throw in a third vector and take that output and check in which direction the movement is going. Still easy and deterministic, so throw in the fog of war (some real fog), a PIR does not work very well in fog so results soddenly become vague and random. We so not want the sentry to go firing off its guns at random so we need to feed each of these results into a hidden layer that will mash them all together in some way that gives the results that best match practical experience. This puts a weighting on each of the various inputs and combines them to give an output, the weighting is found by training the neural net against many thousands of different atmospheric conditions and taking the results that fit our preconceptions of a most likely outcome. This was a simple example in a real world case there might be a fairly large number of hidden layers each with different weighting factors.

If you now try and figure out just why the ‘sentry’ fired its gun and killed your friend Joe you will get a ‘probably’ because of this and that, but the reality is that it is a ‘beats the crap out of me and, I can only give you a best guess as an answer’.

Its a bit like showing someone a shade of grey and asking if it is black or white. The question is inappropriate as the answer will be shaded by what do you mean by black or white, and all the external factors such as lighting etc.

Currently neural networks cannot give an understandable answer for why it comes up with a particular action. You have to trust that the programmer did not slide in a few biases (for example set that everything above the size of a small dog may be a viable target i.e small children are viable. You can also throw in biases during the classification of the many sets of training data.

In other words you have to have complete trust in a programmer, who if honest will still give a ‘beats the crap out of me’ response when asked ‘Why?’

I’ve horribly simplified my answer, a better response might be had by looking at this. (ignore the programming and concentrate on the pics and headings to possibly get a better over-view.)

It’s self learning in the way of the more and more photos that goes into the system then greater sample set it has, so the better it’s guess is.

The same as we self learn by reading. What we don’t have yet is self thinking. I do think Google ‘ai’ does learn from wrong predictions, but just like humans it’s needs to be told it’s choice wrong.

Alot of the capcha cards “pic all the photos with hills in” for eg. And Google has its own set up to quicken this called iirc  Google rewards, it wants people help in identifying similar pictures, like bears v dogs from differnt angles, and it makes games out of drawing games, ie seeing have many people draw the same objects. A bit like global pictionary.

Also the AI should not kill … Isn’t from toms hardware it is one of Asimovs laws of robotics. Form about 1930. I think they borrowed it.

Speedy the most dangerous time with grieving will be when we have AI and humans driving together. I love driving but we are the weak link. An AI can’t calculate for our stupid human decisions. In theory if all cars can talk to each other, and know what each each is doing, we could all drive nose to tale at 200mph up the M1, in perfect safety. Though motorways ate the simple bit.

There is more and more smarts going into cars, and I can see it’s all going towards giving us some choice, in the medium term in if we want autonomous or not. However year on year, control will be taken off us bit by bit. To the point where the autonomous cars will get to override the non autonomous ones. Ie, you’ll go to make a lane change and your car will step in and say “not yet”. It’s only one step away from blind spot detection and lane keep, we already have.

I really doubt Google’s cloud is self-learning image recognition

Don’t doubt. I already have access to this technology for CCTV and it doesn’t cost the earth. You need a Hikvision 4 line camera and a DeepInMind NVR for Deep Learning Analysis. The processing power of the NVR would normally be capable of servicing 32 cameras but can only handle 4 for this type face recognition.

With 4TB the NVR is £5k, a typical camera is £300 (inc vat). Rather than me ramble on watch this live demo video from my supplier https://www.youtube.com/watch?v=8upp8w_voPw

Amazing. For those who didn’t bother you can add a face to a database and trigger alerts when they are detected. I’ve done my hands on training on this now and one thing you could do is detect a know shop lifter, swing a PTZ around and have it follow them whilst alerting the mobile phone of the store security.

Another thing is the search can access multiple NVRs, say on a large retail site. Lost child? You can bet mother has a photo on her somewhere. Scan that in, search and all the hits will appear on a map with time stamps.

If I can do this imagine what Google is capable of.

Isn’t that what humans do every day ed. Best guess? Our brain is much like a cpu, when we go to take over another car, we quickly without knowing it, do some mighty complex equations, on closing speeds distances, and time etc then we use our gut/experiance to decide wether we should go.

We as a whole, best guess all day long, based on learnings others gave us, and own experiance.

Trading has been an issue for years, speed is kind. Billions is spent just on high speed links. AI or rather computer analytics is perfect for stock markets.

I can see a near future where a pc runs a central bank or2. Once the second jumps on the wagon, they will all jump on board.

Just to throw another spanner at the dystopian future, Quartz muses that AI could be used to judge morality. Just imagine coupling that up with a mechanical RoboCop — ‘Excuse me sir, it looks like you have offended MY moral code, your penalty is immediate termination!’.

I think Steve was implying ‘self directing’ in that currently AI only acts within set parameters. If ever AI systems link up and become self directing then they could be truly God-like, not something I think humanity could tolerate though Musk seems to think we are on a slippery path to that fate. Vanity Fair Article

There is a school of thought that says all learning is based on desire so as long as we are still in charge of setting the desired resault we should be safe to let AI figure out the best way to get there.