Creating AI using biological network techniques
- 1. CREATING AI USING BIOLOGICAL NETWORK TECHNIQUES Dr Janet Bastiman, Chief Science Officer Story-Stream.com STORY-STREAM.COM
- 2. “Human subtlety will never devise an invention more beautiful, more simple or more direct than does nature” STORY-STREAM.COM STORY-STREAM.COM
- 3. - Neurons have memory - Connections are dynamic - Computational models are simplistic Integrate and fire is convenient but easy. NEURONS ARE NOT TRANSISTORS
- 4. Multiple connections of different types gives complexity and adaptability. Circuit break at synapses gives some interesting dynamics. BIOLOGICAL CONNECTIONS ARE COMPLEX
- 5. SYNAPSES PROVIDE CONTROL Different synapses are affected in different ways. • Signal fatigue/strengthening • Amplitude and frequency of action potential • External influence Frequency of firing effects the signal strength.
- 6. Still feed forward Memory within layers to impact future events Does very well on predictions Could we do better? LONG SHORT TERM MEMORY NETWORKS, IS THIS ENOUGH?
- 7. NEURONS ARE PACKED DENSELY This is not a simple feed forward model. • Multiple pathways can affect each other • Diffusion can alter neurons without direct connection • Circular connectivity in places Neurons exist in 3D space with major interconnectivity.
- 8. ADDING SPATIAL COORDINATES Adding location allows diffusion to connect layers differently Responsiveness to diffusion can be trained to achieve its own levels of activation and fatigue • Level of detail depends on problem • Overhead to reward not suitable for all problems Start simple – add complexity as needed. Affect disparate parts of your network at run time.
- 9. General intelligence is hard Some specific tasks are also difficult with existing techniques “More data” not always possible Biology doesn’t need large amounts of data so what can we learn Why make networks more complex than they already are?
- 10. STORY-STREAM.COM STORY-STREAM.COM NOBODY WANTS AN EMOTIONAL CAR, OR DO WE?