Thursday, March 22, 2018

AI researchers, wake up!

Picture from Jeff Bezos' Tweet
"Taking my new dog for a walk"
I wanted to write this post since a while but have been putting it off due to other priorities but with the recent whistle-blowing about Cambridge Analytica and also my MVP friend James Ashley starting an AI Ethics blogging initiative, it's about time.

I would say I'm fairly knowledgeable in the AI field staying up-to-date with the research field and currently exploring WinML in Mixed Reality. I also developed my own Deep Learning engine aka Deep Neural Network (DNN) from scratch 15 years ago and applied for OCR. It was just a simple Multi-Layer Perceptron but hey we used a Genetic training algorithm which also seems to have a renaissance in the Deep Learning field as Deep Neuroevolution.

If you are familiar with Deep Learning skip this paragraph but if not: A DNN is simply put simulating the human brain with neurons connected with synapses. Artificial neural networks are basically lots of matrix computations where the trained data is stored as synapse weight vectors and tuned via lots of parameters like neuron activations functions, network structure, etc.
Those networks are trained via different ways but most common these days is supervised learning where lots of big data training sets are run through the DNN, a desired output is compared with the actual output and the error is then backpropagated until the actual output is sufficient. Once the DNN is trained it enters the inference phase where it is provided with unknown data sets and if it's trained correctly it will generalize and provide the right output on unknown inputs.
The basic techniques are quite old and were gathering dust but since a few years there is this renaissance when lots of training data became available, huge computing power in form of powerful GPUs and new specialized accelerators, plus AI researchers discovered GitHub and open source.


AI researchers, the question has to be: Should we do it, not can we do it.

We live in the era of Neural Networks AI and it only has just begun. Current AI systems are very specific and targeted at certain tasks but Artificial General Intelligence (AGI) is becoming more and more interesting with advances in Reinforcement Learning (RL) like Google's DeepMind AlphaGo challenge.

RL is just in its infancy but already quite scary if you look at some research from game development innovators like SEED literally using military training camp simulations to train self-learning agents.


They also applied it to Battlefield 1. There are a couple of things to improve but nevertheless impressive results were achieved which are at beyond AlphaGo.
Does everyone see an army of AI in this video or is it just me? 



I'm sure the involved developers don't have bad goals in mind and I can see it being nicely suited for computer games AI but it's not just the age of Deep Learning but also real-time ray tracing is moving forward, so why not render photorealistic scenes and use that as input for your RL agent training which is then deployed to a real-world war machine. With the photorealistic quality of ray tracing endless real-world-like scenes can be synthesized solving the problem of getting enough training data.
Below is another video showing real-time ray tracing and self-learning agents. The video looks cute and all fun but think about if you replace the assets with a different scene.
Don't get me wrong it's super impressive and I'm all in for advancements in tech but we have to think about the implications in a broader context.


Oh wait, there's actually already a photorealistic simulation framework available to train autonomous drones and other vehicles, they just need to add real-time ray tracing now but I guess it's already in the works. I was getting a little worried but good that drones are only used in civilian scenarios.



Now look at the recent Boston Dynamics robots like Atlas or the SpotMini robot dog which inspired Black Mirror for a good reason.


But hey isn't it cute how Jeff Bezos is just taking his robot dog out for a walk?

We don't even have to look at the future of RL and AGI and scary robots with the fear of the singularity, we already have quite amazing achievements especially with LSTM and CNN type neural networks. Some of them outperform humans already. LSTMs are used for time-based information like speech recognition and synthesis and CNNs for computer vision tasks. 

The author of xkcd shares the same concerns













The interesting and scary part is that artificial neural networks are almost a black box after being trained and can make implications the developer is not aware of. There's always that uncertainty with DNNs, at least for now and it can have huge implications for racial, gender and religious profiling even if that was not the intend of the developer/researcher.

AI researchers, wake up.

Think twice before working on the next cool thing that raises your reputation in the research community. There's more to it than your research and work silo. It's the whole humanity. You have great power and with that comes even greater responsibility.
Ask yourself constantly: What are the implications? Should we do it?


Let's look at some more amazing examples and their implications 

I could post tons of recent examples which are super cool if you wear the geek hat but which are super scary if you put on the ethics hat on and think a bit about their implications.

Autonomous vehicles are the future and I'm sure lots of us can't wait until we can relax during a long drive in our self-driving car but it might not be ready yet for tests out in the wild like Uber's deadly accident shows. No doubt that human drivers are worse even today but the perception is that AI is much better and it should be indeed. My guess is there's likely not enough training data available for all the edge cases yet since the object movement detection should have triggered a full stop but it might have been interpreted as a moving light shadow but the radar sensor should get it. We will see what the investigations find out but I still think it's too early for these kind of real-world tests.

NVIDIA has some amazing research ongoing with unsupervised mage-to-image translation. Just watch this demo video below and then think about if we can be sure if dashcam footage was really recorded during day or night.



Google's WaveNet and even more so Baidu's DeepVoice show impressive results for speech synthesis using samples of humans and then synthesizing their voice patterns. The amount of sample data needed to fake a person's voice is getting less and less, so not just public figures with lots of open samples but basically everyone can be imitated using text-to-speech.

It doesn't stop with audio synthesis. Researches from University of Washington made great progress with video synthesis. Play the embedded video and think about the implications this tech could have being in the wrong hands.


You might have heard about Deep Fakes videos being mainly used to generate fake celebrity porn but even worse things like this were created.

Well and of course video surveillance is getting a huge boosts since a couple of years due to great progress in face and object tracking Deep Learning AI. BBC tested a system in a Chinese city which is fully covered with AI-enabled CCTV cameras and it did not disappoint.
Make your own conclusion if it's a good or a scary thing that everyone is always tracked.


There's light!

It's not all just dark, there are of course as many good examples available that leverage modern AI for a good cause with little dark implications.

Impressive results were achieved for AI lipreading that beats professional, human lip readers which can help many people to live a better life.


Also huge advances are achieved in the medical field and especially in the computer vision tasks to automatically analyze radiology images like breast cancer mammography or improving noisy MRI data.

Also prominent companies like Google's DeepMind are beginning to realize the implications for humanity of their work and have started ethical initiatives.

And then there's Facebook!

It's crazy how ethical questions play a little role for large companies like Facebook which is hoarding billions of data sets from around the world which can be used for training. We even provide them not just the input data sets but even the training output with our likes, clicks but even just the scrolling behavior when you read. Plus the huge investments in Facebook's AI Research group hiring and growing some of the best AI talent in the industry.
Just look at some of the research areas of FB like DeepText which is super impressive and aims for "Better understanding people's interest". Now ask yourself for what? Ads? What kind of ads? What is an ad? Is a behavioral changing FB feed an ad?
And then you have companies like Cambridge Analytica who crawled/acquired the data and abuse it to sell their information warfare mercenary services to anyone changing human behavior and altering elections.
Real-world war machines might not be needed anymore, Big data + Deep Learning + Behavioral Psychology is a dangerous weapon if not the most dangerous.
It's good to see Mark Zuckerberg apologizing for the issues and we can only hope it will have real consequences and is actually still controllable at all.

And then there's YOU!

It's should not just be the Elon Musk's of the world warning about the impact of unethical AI, we as developers and researchers being at the forefront of technology have a responsibility too and need to speak up. It's about time, so please think about it and share your thoughts and raise your concerns.

The creator of Keras, a popular Deep Learning framework, a real expert shared his thoughts about Facebook and the implications of its massive AI research investments.
I couldn't agree more so let me finish this post here with his must-read Twitter thread:

Make sure to read the whole thread here.
There's a little twist to it since François works for Google but I expect he sticks to his own principles.

AI researchers, wake up! Say NO to unethical AI! 


No comments:

Post a Comment