Ankur Narang; Senior Vice President – Technology & Decision Sciences at Yatra at Cypher2017- Part1/3
summary
Claims that Deep Reinforcement Learning (DRL) combines convolutional and recurrent nets to learn policies from pixels rather than hand‑crafted features, with concrete results: Atari DQN beating humans on specific games, AlphaGo achieving superhuman Go play, and Udacity contestants using 3D convolution + RNNs (sequence length 10) to predict steering, speed and torque from video. Examples of tools include DQNs, policy gradients and LSTMs; image QA reaches ~73% on some sets and video captioning still shows frequent semantic errors, highlighting sample complexity and generalization gaps.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 1:05Deep Learning
- 7:10Neural Networks
- 14:30Machine Translation
- 17:13Video Description Generation
- 19:38Safe Driving Cars
- 20:51Reinforcement Learning
related talks
28mDeep Reinforcement Learning: theory, intuition, code
01h 17mTraining intelligent game agents using deep reinforcement learning
01h 03mIEEE CoG 2019 - Industry Day - David Silver (DeepMind)
33mBreakthroughs and the Future of (Deep) Reinforcement Learning
49mReinforcement Learning: Demystifying the hype
05mPractical insights into deep reinforcement learning
44mReinforcement learning: A gentle introduction
49mML: Teaching machine to learn from its own mistakes (...)