Context #
This article contains most of the content of the slides of the presentation by Demis Hassabis available at https://www.youtube.com/watch?v=IjG_Fx3D0o0.
Learned in this study #
Things to explore #
Overview #
- How can we know/measure we’re making progress toward AGI
- Non-biological approach vs biological approach
- Issues with the non-biological approach
- Brittle
- Time-consuming to train
- Poor at general learning
- Difficult to acquire/generate new symbols
- How do you refer to things outside of the agent? (symbol grounding problem)
- Biological approach
- The brain as a blueprint
- Covers a large class of approaches
- Different search spaces of possible AGI solutions
- Regime 1: Small and dense search space
- Not worth too much relying on the human brain design
- Regime 2: Large and sparse search space
- Worth a lot to rely on the human brain design
- Regime 1: Small and dense search space
- Evidence points to regime 2:
- Evolution has only produced human level intelligence once
- Large non-biological projects failed to make progress
Approaches to AGI (from abstract to biological) #
- Cognitive science architectures: SOAR (Laird/Newell), ACT-R (Andersen), OpenCog (Goertzel)
- Unsatisfactory because they’re besed on introspection and when changes in knowledge occurs, they have to modify their model to fit in this new understanding
- System neuroscience: the brain algorithms
- Brain emulation: Blue Brain (Markram), SyNAPSE (Modha)
- Not telling us about the internal processes/functions going inside the brain
- Relying on very intricate imaging techniques (at what level do we need to stop? Calcium ion channels? Atoms?)
Marr’s three levels of analysis #
- Computational: What - the goals of the system
- Algorithmic: How - the representations and algorithms
- Implementation: Medium - the physical realisation of the system
Rapid advances in neuroscience #
- Revolution in cognitive neuroscience
- New experimental techniques
- Sophisticated analysis tools
- Exponential growth in understanding
- Actively conduct neuroscience research useful for building AGI
Role of neuroscience #
- Likely that neuroscience will have a big role in building AGI
- As an orthogonal source of information to Machine Learning
- Provides direction: inspiration for new algorithms/architectures
- Validation testing: does an algorithm consistute a viable component of an AGI system?
- How can it not be a net benefit in the quest for AGI systems to add neuroscience knowledge into the mix?
The hybrid approach #
- Combine the best of machine learning and neuroscience
- Where we know how to build a component
- Use the latest state-of-the-art algorithms
- Where we don’t know how to build a component
- Continue to push pure machine learning approaches hard
- In parallel, also look to systems neuroscience for solutions
Systems neuroscience procedure #
- Extract the principles behind an algorithm the brain uses
- Creatively re-implment that in a computational model
- Result: a state-of-the-art technique and AGI component
Intermediate goals #
- Full embodied physical robots: throws up complex engineering problems whilst distracting from the main problem of intelligence
- Toddler AGI: AI-controlled robot that display qualitatively similar cognitive behaviours to a young human child (~3yo)
- Massive breadth of capabilities required = extremely hard
Core AGI #
- Core capabilities:
- Conceptual knowledge acquisition/representation
- Planning and prediction abilities
Concepts are key #
- Knowledge in the brain
- Symbols
- Conceptual
- Perceptual
- Equivalent machine learning algorithms
- Logic networks
- ???
- DBN, HMAX, HTM
- So how does the brain acquire conceptual knowledge?
Hippocampal-neocortical consolidation #
- Hippocampus sits at the apex of the sensory cortex
- High-level neocortex: association and prefrontal cortex
- Stores the memories of recent experiences or episodes
- Replays those memories during sleep at speeded rates
- Gives high-level neocortex samples to learn from
- Memories selected stochastically for replay
- Rewarded: emotional or salient memories replayed more
- Circumvents the statistics of the external environment
- (Hypothesis) Leads to abstraction and semantic knowledge
Interim milestones #
- Build knowledge on top of existing knowledge
- Abstract classification: classification of empty/full containers
- Discovery of higher-order structures (eg. 123456789101112131…) What is the next number? Statistics is not enough
- Algorithms that can build sophisticated models of the environment (eg. play any card game just by observing a raw perceptual stream)
- Transfer learning: learning a response in one perceptual context, abstracting a rule, and applying it correctly in a new context
- Some impressive things have already happened:
- MoGo - first program to beat a professional human go player
- IBM’s Watson - taking on human champions at Jeopardy quiz show
How to measure progress? #
- One approach: measure success across a suite of tasks
- Ideally we’d like a more integrated measure of progress
- Algorithmic Intelligence Quotient (AIQ)
Predictions #
- Systems neuroscience understanding will help inspire to several key components of the overall AGI puzzle
- System with transfer learning and conceptual knowledge acquisition capabilities will appear in the next 5 years
- Measurement tools charting progress are improving all the time
- Once interim milestones have been achieved, we will have a better understanding of of intelligence and the safety issues involved
- Probably ~20+ years for full human-level AGI but lots of interesting technologies will be built on the way
See also #
- Marr’s tree levels of analysis:
https://en.wikipedia.org/wiki/David_Marr_(neuroscientist)
- Algorithmic Intelligence Quotient:
http://arxiv.org/pdf/1109.5951.pdf