newsfilter.io
Lecture, Keynote

Deep Learning State of the Art (2020)

  • Predicts 2020 as a year of significant activity and progress following previous years, anticipating a shift in the popular press to declare the end of the deep learning era.
  • Envisions a community environment by 2020 characterized by reduced hype, criticism, derision, and jealousy, replaced by solid research, respect, open-mindedness, and collaboration.
  • Anticipates increased research output in reasoning, common sense, active learning, lifelong learning, multimodal/multitask learning, open domain conversation, and algorithmic ethics including fairness, privacy, and bias.
  • Expects continued advancements in deep reinforcement learning applications, robotics, and specifically legged robotics and robot manipulation.
  • Forecasts TensorFlow 2.0 and PyTorch 1.3 converging to combine strengths while eliminating weaknesses, with support for Python 2 ending on January 1st, 2020.
  • Hopes for research to become framework-agnostic to facilitate easy model transfer between TensorFlow and PyTorch.
  • Expects OpenAI and DeepMind to advance reinforcement learning frameworks to maturity levels comparable to OpenAI Gym by 2020.
  • Predicts greater abstractions enabling scientists outside machine learning to utilize deep learning without extensive Python programming knowledge.
  • Anticipates reasoning and common sense reasoning becoming central components of transformer-based language model work.
  • Expects transformer context windows to expand from hundreds or thousands of words to tens of thousands, allowing for the processing of entire stories.
  • Predicts the success of transformer architectures transferring to visual information and video analysis.
  • Expects reinforcement learning to be utilized for exploring social behaviors in agents and their echo in human systems.
  • Envisions applied deep learning innovation in autonomous vehicles focusing on active, multitask, lifelong, and online learning.
  • Forecasts increased over-the-air updates for autonomous vehicles as a prerequisite for solving autonomy problems.
  • Hopes for more public datasets of edge cases released by automotive companies alongside simulators like Carla and NVIDIA Drive Constellation.
  • Expects reduced fear of AI and enhanced government-expert discourse regarding privacy, cybersecurity, and transparency.
  • Identifies recommendation systems as the most exciting and powerful AI space for the next couple of decades due to societal impact.
  • Anticipates breakthroughs in open domain conversation within the Alexa Prize, despite current estimates suggesting such achievement is two or three decades away.
  • Predicts that within approximately 20 years, society will attribute rights to robots similar to those currently afforded to animals.
  • Expects Artificial General Intelligence (AGI) systems to exist as companions and digital assistants rather than posing a threat to human life.
  • Predicts AI systems will never become human masters, with the primary danger lying in tech company owners using these tools for control.