Highlights of AI Frontiers Conference (Nov 9–11, 2018)

By Mohan Reddy, CTO of The Hive

Although we live in exciting times, at times it feels like the “Irrational Exuberance” of AI hype and adoption. AI Frontiers conference was a welcome change in that respect wherein the best brains of AI, industry practitioners and leaders came to talk about applications, use cases and cutting-edge research. AI Frontiers Conference was at the San Jose Convention Center and was very well attended.

Big Ideas

Ilya Sutskever was the opening keynote speaker and talked about OpenAI’s mission, why near-term AGI should be taken as a serious possibility and very much can:

  • Generate massive wealth
  • Potential to end poverty, achieve material abundance
  • Generate science and technology
  • Cure disease, extend life, superhuman healthcare
  • Mitigate global warming, clean the oceans, etc.
  • Massively improve education and psychological well being
“OpenAI’s mission is to ensure that artificial general intelligence (AGI) — by which we mean highly autonomous systems that outperform humans at most economically valuable work — benefits all of humanity.”
— The OpenAI Charter

According to Ilya, whether you find this evidence compelling or not is less important than the framework: as a community, let’s rely less on intuition and more on hypotheses and evidence to guide our conversations about AGI. We are building technologies that we all feel are going to have a massive impact on the world, and it’s on us to be good custodians of that responsibility.

The talk also included some big ideas on very large scale reinforcement learning, self play, and reward shaping.

Fig: OpenAI trained a human-like robot hand to manipulate physical objects with unprecedented dexterity.
Fig: Domain randomization, learns in simulation which is designed to provide a variety of experiences rather than maximizing realism.

Ashok Shrivastava, Chief Data Officer of Intuit, talked about using AI to solve complex economic problems. Small businesses are the very fabric of the world economy and the U.S. economy — and Intuit uses AI and ML to help small businesses take advantage of the same capabilities that very large companies do. AI/ML automates financial tracking in multiple Intuit products. One big challenge business owners face is getting a better understanding of their finances. Intuit looks to democratize financial literacy, using AI and ML to help QuickBooks users get the insights they need more efficiently. 13.4 million QuickBooks and Mint customers can automatically categorize 250B transactions per year with >90% accuracy.

Fig: The AI flywheel for finance at Intuit.

Melissa Goldman, MD and CIO at JP Morgan Chase, spoke about how financial institutions have an opportunity to tackle significant socio-economic challenges using AI/ML.

Dr. Kai-Fu Lee talked about how AI is entering the era of implementation and that data plays an important role. When it comes to data China is a leader! Better products lead to more users, those users lead to more data, and that data leads to even better products, and thus more users and more data. Chinese and U.S. companies are leaping out to massive leads over the rest of the world.

Deep Learning Breakthroughs

Designing neural network architectures is a difficult task and lots of human effort goes into tuning them. So there is a need to try and learn “good architectures” automatically. AutoML aims at automating AI. It uses RNN as a controller to generate Child Network and uses reinforcement learning to update the Controller based on the accuracy of the Child Network. Teaching machines how to learn is a new way to do computer programming: instead of writing the program, you teach a machine to do it said Quoc Le, the inventor of AutoML.

Data Wrangling is a process of transforming the data from its raw format to a more structured format that is amenable to analysis and visualization. It is estimated that data scientists spend 80% of their time wrangling data. As software evolves, developers edit program source code to add features, fix bugs, or refactor it for readability, modularity, or performance improvements. The two killer applications for ‘programming by examples’ today are in the space of data transformations/wrangling and code transformations. Learning to learn programming by examples is the next frontier in this direction letting end users construct and run new programs by providing examples of the intended program behavior.

Cutting edge work

Anima Anandkumar, Director of ML research at NVIDIA, talked about Tensors as learning representations that can encode data dimensions, modalities and higher-order relationships. Embarrassingly parallel Tensor operations have inspired several compact and more accurate neural network architectures.

Robotics

According to Pieter Abbeel, Professor at UC Berkeley, there are two waves of automation — 1. Robots with “eyes, which is happening now, and 2. Teachable Robots (“get help anywhere, anytime”). There are many exciting challenges for AI in Robotics such as:

  • Fast Reinforcement Learning
  • Fast Imitation Learning
  • Leverage Simulation
  • Model-based RL
  • Long Horizon / Hierarchical Reasoning
  • Safe and Lifelong Learning
  • Value Alignment

Navin Budhiraja, VP and GM at ANSYS Inc., talks about how simulation combined with AI will significantly augment product design, manufacturing processes. Generative design powered by AI will optimize new products for improved manufacturability, reduce costs and decrease manufacturing cycle times.

Fig: Navin Budhiraja on how AI being used in design.

Conclusion

We are going to see AI now moving into the era of applications such as AI-assisted product design, IoT, enterprise applications, etc. There will be more emphasis on data labeling and model explainability. This concert of technologies discussed above will lay the foundation to combine with other emerging technologies to create breakthrough opportunities.