Artificial general intelligence research project at Keen Software House (3/2015)

  • Published on

  • View

  • Download


Artificial intelligence research at Keen Software House Artificial intelligence research at Keen Software House Marek Rosa, Dušan Fedorčák, Martin Bálek Keen Software House March, 2015 Introduction About us Project history How does our team work Why general artificial intelligence? Long-term goal Short-term goals: R&D & commercial Open problems Second part (more technical), brain simulator architecture – Dušan and Martin About us Interest in AI and robotics since age of 15 How to achieve it? Space Engineers Medieval Engineers *** Engineers  Set up an AI team Shared common goal Actual R&D funds: $10mil Project history Started in 1/2014 Examining various AI approaches from biologically based neurons (e.g. spiking) to very artificial solutions Brain Simulator Milestones: Pong Team grown to 12 researchers and the plan is 30 or more… How does our team work Milestones vs. autonomous research to pursue creative solutions 2 team meetings each week (update, brainstorming) => knowledge sharing Rapid iterations => fail fast, fail often, fail forward Studying and experimenting Motivation: working on the most exciting scientific challenge = meaningful work What’s next: Early access Openness Ecosystem External pressure Why general artificial intelligence? Narrow vs. general AI Highest “return on investment” (ROI) possible => high-risk & high-reward Recursively self-improving AI Exponential growth Market size: unlimited Could be “our final invention” (in a good sense) AI scientists, AI programmers, AI astronauts, AI *** Next step in evolution AI will change everything Everyone will benefit from AI (charities, corporations, individuals…) The future will be awesome! Credit: "The Singularity is Near" Long-term goal Long-term goal: human-level AI in 10 to 50 years What is general AI? Artificial brain that can perceive, learn and adapt to the environment while maximizing its short and long term rewards Sensors Input: visual, auditory, tactile, etc. Motors Output: e.g. sequence of muscle commands Motivations Input: reward and punishment Brain: architecture of AI modules that learn the patterns and sequences of signal coming in and out of the brain; also patterns within the brain spatial and temporal seeking causalities and correlations finding associations working memory prediction for modules that can benefit from seeing the future long term memory goal selection and hierarchical goal execution (based on motivations) all this on multiple levels of hierarchy and more: feature extraction, generalization, abstraction, etc. Architecture: heterogeneous Learning Not hardcoded Online learning Learns by interacting with the environment and with itself – like children Learning from a mentor (mirroring). Doesn’t need to waste time by exploring solutions that won’t lead to useful outcomes. Brain Sensors Motivations Motors R&D short-term goals Already accomplished: AI that learns to play Pong Unstructured input (screen pixels and reward/punishment signals) AI has to extract useful features from the image, causalities, correlations, select goals that lead to increasing reward and avoiding punishment Google DeepMind Upcoming milestones: AI that plays a game with a more complex environment; delayed reward that requires long-term goal following AI that learns to play variety of games without the requirement to “restart the brain” Muscle control sequences, balancing Gameboy Pong Commercial short-term goals AI company AI platform/ecosystem Brain Simulator Marketplace AI module developers AI brain architectures Licensing to customers (robotic firms, AI app developers) Investing in AI developers Community feedback Equity crowd-funding Our basic AI R&D will continue in parallel Open problems AI safety => friendly AI and collaboration Robots will take our jobs! => invest in AI Many problems are unsolvable by narrow AI and require human intuition and knowledge (acquired from birth to adulthood) Can AI fast-forward this process? What if our future human-level AI requires extreme computational resources? (out of our reach). E.g. simulating 100 billion biological neurons Moore’s law is on our side Better start the project today and hope that in 10+ years hardware will be ready Maybe our implementation will use resources better than the nature  What you can do for yourself? 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 me: