At Rise of AI 2017 in Berlin, Damian Borth, Director of the Deep Learning Competence Center at DFKI, traces how deep learning became the engine of the current AI wave. He opens with political context, from Angela Merkel's 2013 remark that the internet is uncharted territory to Barack Obama's 2016 comment that his successor would govern a society transformed by AI, and to AlphaGo's 2016 win over Lee Sedol. He explains why 2012 was the turning point for deep learning, when AlexNet won the ImageNet competition by a large margin thanks to three factors coming together: deeper networks, the large labeled ImageNet dataset, and GPU compute that Nvidia's Jensen Huang chose to support once researchers started repurposing graphics cards for training. Borth then covers post-2012 architectures, the privacy risks of models that memorize training data, his own team's pioneering work extracting sentiment from visual content, and the resulting Nvidia AI Lab recognition. He closes by arguing that society, not companies or politicians alone, must decide where AI's use should stop, and calls for risk-based classification and compliance systems for AI.
Key takeaways
- Borth identifies three factors converging in 2012 as the reason deep learning took off: deeper network capacity, the large labeled ImageNet dataset, and GPU compute.
- Nvidia's decision to support researchers who were repurposing graphics cards for neural network training now generates about 1.5 of Nvidia's roughly 5 billion dollars in revenue, according to Borth.
- Borth warns that because trained models compress their training data, they can pose privacy risks, and urges companies to treat their data as their own intellectual property rather than handing it to outside AI providers.
- Borth's team was, by his account, the first worldwide to extract sentiment from visual content in real time, work that earned his group one of only two Nvidia AI Labs in Europe at the time.
- Borth argues AI systems should be classified by real-world impact, with regulation focused on high-stakes uses like self-driving cars rather than low-stakes classification tasks.
Detailed notes
From Merkel's 'Neuland' to AlphaGo
Borth opens with political framing: Angela Merkel's 2013 remark that the internet was uncharted territory, made shortly after the Snowden revelations, and Barack Obama's 2016 comment that his successor would govern a country being transformed by AI, a remark Borth notes Obama made without knowing Donald Trump would be that successor. He then recounts AlphaGo's March 2016 match against world champion Lee Sedol, in which the machine won three of the first three games, lost the fourth, and won the fifth, and mentions an internet meme, which he calls an early example of fake news, misreading the board's stone pattern in that fourth game. He adds a recent quote he attributes to a political figure stating that a country not keeping up with deep learning and neural networks will lose.
Why 2012 was the turning point
Borth explains that neural networks, which mathematically simulate connected neurons to map inputs to outputs, are an old idea that only became powerful in 2012 when three factors converged: growing network capacity through more neurons and layers, the large labeled ImageNet dataset needed to fit those extra parameters, and GPU hardware able to do the necessary computation. That year, Alex Krizhevsky's eight-layer AlexNet won the ImageNet challenge by a wide margin over all other methods. Borth notes that researchers had been 'misusing' graphics cards for neural network training, and that Nvidia's CEO chose to support rather than block this use, a decision Borth says now generates about 1.5 billion of Nvidia's roughly 5 billion dollars in revenue from GPUs that never power a monitor.
Opening the black box
After 2012, open-sourced frameworks such as Caffe let many researchers build deeper architectures, from VGGNet and GoogLeNet's Inception design to Microsoft's 2015 ResNet with residual layers, pushing networks to hundreds or thousands of layers with no sign of saturating performance gains. Visualizing what these layers detect shows early layers acting as generic edge detectors while deeper layers specialize into concepts such as faces, cars or chairs, which enables transfer learning by reusing a trained network's early layers for new tasks. Google's 2015 DeepDream work, which visualizes what a network 'sees,' illustrates for Borth a privacy risk: because models compress the data they were trained on, it can become possible to partially reconstruct a specific person's face from a model, which is why he tells companies, including German car manufacturers, that their data is their intellectual property and should not be handed to outside providers. He describes an emerging shift from an app-store paradigm to a 'model store' paradigm, where trained models themselves become tradeable assets, illustrated by neural style-transfer work from the University of Tuebingen that fuses a photo with an artistic style.
Borth's own work: sentiment from visual content
Borth describes his team, working with UC Berkeley and Columbia University, as the first worldwide to extract sentiment from visual content, running in real time on video to identify adjective-noun concept pairs with a positive or negative sentiment score. He cites potential uses in targeted advertising from surveillance footage and in mapping a city, such as Berlin, by perceived safety or desirability, work he says interests a real estate company managing 60 billion dollars in assets that wants to assess distant markets like India or China without sending people there. This work led Nvidia to award his group one of only two Nvidia AI Labs in Europe. He also describes AlexNet's training time falling from two months in 2012 to two weeks two years later and to two hours with 2016's Pascal GPUs, and notes that Nvidia's announcement of its next-generation Volta GPUs added about 10 billion dollars to its market capitalization in a single day.
AI's economic scale and Borth's call for responsibility
Borth estimates the AI market at about 500 billion dollars, with 5 billion dollars in VC funding already invested, disrupting industries from automotive and retail to finance and pharmaceuticals, and warns researchers to manage expectations rather than overpromise. He argues society, not a single company executive or politician, should decide where AI's application must stop, and calls for classifying AI systems by their impact, focusing regulation on higher-stakes systems such as self-driving cars or robotic surgery rather than low-stakes ones like classifying cats and dogs, alongside future AI compliance and AI law-enforcement roles. He mentions creating Germany's first free online machine learning course, launched at CeBIT with about 5,000 participants and 338 video clips featuring contributors from Google, Nvidia and Amazon. He closes by invoking Adobe co-founders Chuck Geschke and John Warnock, who said they were aware of their responsibility while building the desktop publishing industry, and asks whether today's AI community is equally aware of its responsibility.
Quotes
It's a non-convex optimization problem, which means try and error.
Damian Borth
As a company, be aware that your data is your IP, and don't give your data away.
Damian Borth
We are today in the middle of a revolution with AI, and I'm wondering who is aware of the responsibility that is actually actively modifying and pushing this area.
Damian Borth
Questions answered in this talk
Why was 2012 a turning point for deep learning?
According to Damian Borth at Rise of AI 2017, 2012 was when three factors converged: neural networks with enough layers to have real learning capacity, the large labeled ImageNet dataset needed to train them, and GPU hardware capable of the required computation, exemplified by AlexNet's decisive ImageNet win.
Why does Damian Borth warn companies about giving their data to AI providers?
According to Damian Borth at Rise of AI 2017, trained neural networks compress the data they learned from, which can make it possible to partially reconstruct information like a specific person's face, so he advises companies, including German car manufacturers, to treat their data as their own intellectual property.
What kind of AI regulation does Damian Borth call for?
According to Damian Borth at Rise of AI 2017, AI systems should first be classified by their real-world impact, with stricter oversight and eventual AI-compliance requirements focused on high-stakes systems like self-driving cars or robotic surgery rather than low-stakes tasks such as classifying cats and dogs.