Numenta's Director of ML Architecture Lawrence Spracklen presented a talk at the SBMT Annual Congress on July 10th, 2021. He talked about how neuroscience principles can inspire better machine learning algorithms.
2. Numenta
Developing machine intelligence
through neocortical theory
• Understand how the brain works
• Apply neocortical principles to AI
Developed the “Thousand Brains”
theory of how the neocortex works
3. Artificial Neural Networks (ANNs)
Layer 1
Layer 2
Layer 3
Layer N
Input
Output Dense, fully-connected and computationally expensive
4. Traditional approach to ANNs
Perform matrix multiplications very fast
• GPUs have become AI workhorses
• 500+ trillion arithmetic operations per
second per card
• Hardware performance doubles
every few years
• Hardware cannot keep pace with
growth in model size
• Exploding AI costs
• 2018 : BERT cost $6K+ to train
• 2020 : GPT-3 cost $10M+ to train
3-years
17,000X
increase
Figure credit
5. AI Today
Incredible progress, but, at what cost….
• Vast models
• Trillions of parameters
• Expensive training
• Massive compute, power & training data requirements
• Catastrophic forgetting
• Static task-specific models that can’t learn
• Fragility
• Significant real-world dangers
Still a long way from AGI (Artificial General Intelligence)
• Can we continue down this current path?
6. Going forward
1. Improve model performance
2. Decrease frequency of retraining
3. Decrease training complexity
• Both Algorithms and Hardware need to evolve
• Focus on just one dimension doesn’t solve the problem
• “Faster Horse” issues
• Ensuring synergy provides a lasting solution
• Hardware feasibility needs to influence algorithm evolution
• And vice versa
7. Can Neuroscience help?
Examine the Neocortex
• Neuron interconnections are sparse
• Neuron activations are sparse
• Neurons are significantly more complex than AI’s point
neuron abstraction
• Humans can learn from very limited examples
Numenta’s Roadmap
8. Make models fast
Sparse models
• Deliver comparable accuracy with up to 20X fewer parameters
• Also leverage activation sparsity for multiplicative benefits
• 100X+ reduction in compute costs
• Hardware needs to be capable of exploiting sparsity
• Efficiently avoid multiplying by the zeros!
• 100X on FPGAs, 20X on CPUs with Numenta’s sparsity
9. Always be learning
Active dendrites
• Point neurons only incorporate proximal synapses
• Small proportion of neuron’s total synapses
• Extend artificial neurons to incorporate distal synapses
• Basal synapses used to modulate neuron behavior
• Applying context signals enables networks to learn multiple
tasks and facilitates online continuous learning
• Primes relevant neurons based on context
• Unsupervised determination of context is critical
image credit
10. Reduce learning repetition
Reference frames
• Training an ANN to recognize even a cup requires many images
• 100s of pictures of cups at different orientations, distances, designs and colors
• Separate problem into two base components
• Invariant representation of object
• Understanding of positional relationship to object
• Create robust position independent representations of objects
• Make observer orientation and distance explicit considerations
• Inspired by human grid cells
• Object independent
• Significantly reduces number of training examples
11. Conclusions
• Continued progress in AI is threatened by exponentially
increasing costs
• Insights from the Neocortex provide critical insights for how to
evolve AI
• Numenta has developed neocortex inspired roadmap to AGI
• Already demonstrated 100X AI model speedups using brain
inspired sparsity
• Working to incorporate continual learning and positionally
invariant representations into AI systems
• Reduce both retraining frequency & number of training examples
• Cumulative benefits reduce AI costs by many orders of
magnitude