Artificial Intelligence Research at Keen Software House
Technical Report
Shortly about AGI
• Artificial General Intelligence
- Autonomous agent
- Able to perceive and change its environment
- Able to remember, reason and plan
- Adaptable and able to learn
- Able to communicate
What to use for AGI?
• Classical AI?- Symbolic architectures- Inference machines, expert systems- Planners and solvers, STRIPS
• Artificial neural networks?- Spiking Networks- FFN, RNN- DeepNets
• Multi agent systems?• All of them!
Suitable tool for experiments
• Rapid model prototyping - Integrate existing model- Create (or recreate) new model
• Model insight- Rich GUI & Visualization possibilities- Model structure view (oriented graph?)- Runtime view & execution control
• Heterogeneous architecture- Connect different models together- Able to use various hardware
• Parallel execution- GPU based solution- Cluster solution
Existing tools & inspiration
• Graph of connected modules- ROS
- Matlab / Simulink
- Maya material editor
- Nengo (Eliasmith)
• Specialized libraries (modules)- Caffe, OpenCV, cuBLAS, cuDNN,
- ROS modules
Our solution – Brain Simulator
• Model structure- Nodes, tasks, memory blocks, worlds
• Model view- Graph view (model structure)- Observers (model data)
• simple, numeric, 3D, custom
• Experiments & debugging - Model parameters exposed to GUI- Adjustable observers- Simulation control
• Parallel computing- CUDA (Intel Phi support in progress)- Multi GPU support
Brain Simulator – screenshots
Brain Simulator – modules
• Implemented modules- Feed-forward nets (FFN, RNN, convolution nets, auto associators)- Self-organizing networks (SOM, GNG, K-means …)- Vector symbolic architectures (HRR, BSC)- Hierarchical temporal memory (spatial & temporal poolers)- Spiking networks & STDP- Computer vision (filters, segmentation, tracking, optical flow)- Hopfield network, SVD, SLAM, PID, Differential evolution and many
others
• Imported modules- Caffe, BLAS, BEPU Physics, Space Engineers, Gameboy emulator
• Planned modules- Deep learning & RBMs, Hierarchical Q-Learning
BS Screenshots – SOM
Development methodology
• Iterative/agile approach- Early implementation and experiments- Separated experiments with mockup parts- Milestone oriented (global model iterations)
• Separated experiments (proofs of concept)- Data representation, memory models, temporal data encoding- Learning strategies, goal inference, action selection- Spatial awareness, visual working memory, navigation- Computer vision
• Milestone examples- 6-legged robot agent (integration test)- Breakout/Pong game (reinforcement learning & vision test)- Autonomous agent game (PacMan, Nethack)
Example 1 – walking robot
• Physical world emulation - Connected to Space Engineers game
- 6-legged robot body
- Runtime visual data processing & body control
• Learning from mentor - Hardwired movements
- Learning body state associated with high level movement commands
- Simple vision to action associations
- Totally supervised system
Video of 6-legged robot
Example 2 – Pong / Breakout
• Pong / Breakout game- From bitmap to buttons- Reinforced learning (reward and punishment)- Image processing towards object tracking- Vector symbolic architecture - Goal states extraction - Action learning & action selection
• Existing solutions- Not Q-learning (DeepMind and others before them)- Modular, engineered system- Better insight (faster learning?), sacrificed flexibility
Pong / Breakout model
Visual Processing
Pong / Breakout model
Pong / Breakout BS inspection
Future work
• Next milestone – 2D egocentric game- Advanced visual working memory- Navigation & inner spatial representation of environment- Environment variables extraction, hierarchical Q-learning- Multiple goals and motivations, goal chaining- Motoric systems (bipedal balancing)
• Future milestones- Same model playing different games - Same model instance playing different games- Motoric systems (command sequences unrolling & execution)
• Computing platform improvements- Brain Simulator release (with remote module repository)- HPC solution- Unix systems release
The end
• You can invest in AI companies• Every $1 invested today will return 1,000,000 times
• Join our team – we are always hiring• AI Programmers / Researchers• SW Engineers / Architects• PR Manager / Evangelist
• Follow us: • http://blog.marekrosa.org/• http:// www.keenswh.com/
Thank you.Questions?