AI Architectures

This page will be heavily extended in the near future. For instance, I will document and contrast the listed systems very soon.

AI Platform Architectures

Airbnb Bighead

Airbnb presented Bighead both at Strata and at Data Council (where they go into a bit more detail into architectures of components, e.g. see Deep Thought at 22:04 and Zipline at 28:01). Here is my summary diagram of the system:


Lyft Flyte’s architecture is described here and its component architecture has a dedicated page. Another illustration can be found here. Below is my attempt to sketch out the overall architecture:

IBM Fabric for Deep Learning (FfDL)

The Fabric for Deep Learning is documented both in its paper (figure 1) and on its Github page. Its architecture looks approximately like this:

This section will be extended over the next weeks. Please refer to the landscape section about AI Platforms in the meantime.

Neural Network Architectures

  • DeepMind’s Alpha Systems: AlphaGo, AlphaGo Zero, Alpha Zero, Alpha Star, AlphaFold
  • OpenAI Codex, CLIP (Github), DALL·E, Five
  • AllenAI Aristo, Mosaic and Prior
  • Facebook BlenderBot (uses poly-encoder transformer architecture – might explain bi-encoder → cross-encoder → poly-encoder)
  • BERT, GPT, ELMo, see Google blog post
  • Distributed RL – A3C, GORILA, APE-X, IMPALA, APPO
  • MANNs – memory nets → end-to-end memory nets → NTMs → DNC
  • Modular neural nets, routing networks, capsule networks, neuro-symbolic concept learner
  • Embeddings word2vec → sentence2vec → doc2vec and graph embeddings like node2vec and specialized embeddings like hotel2vec

Cognitive Architectures

See Comparison of Cognitive Architectures (Wikipedia) and Comparative Table of Cognitive Architectures

Intelligent Agent Design

See agent architectures and TouringMachines distinction into

  • Deliberative Architectures
    • IRMA (Bratman et al.)
    • AUTODRIVE (Wood)
    • Behaviour Hierarchies (Durfee and Montgomery)
    • Agent-Oriented Programming (Shoham)
    • Homer (Vere and Bickmore)
  • Non-deliberative Architectures
    • Subsumption Architecture (Brooks)
    • Situated Automata (Rosenschein and Kaelbling)
    • Pengi (Agre and Chapman)
    • Reactive Action Packages (Firby)
    • Universal Plans (Schoppers)
    • Dynamic Action Selection
  • Hybrid Architectures
    • 3T
    • AuRA
    • Brahms (Agent-Oriented Language, BDI architecture)
    • GAIuS
    • GRL
    • InteRRaP
    • TinyCog
    • TouringMachines

Other Architectures

Biological Inspiration

Components from Brain Computation as Hierarchical Abstraction by Dana H. Ballard:

  • Cortex: Long-Term Memory
  • Basal Ganglia: Program Sequencing
  • Thalamus: I/O
  • Hippocampus: Program Modifications
  • Amygdala: Focus

Additional Considerations

  • Different neurotransmitters beyond dopamine
  • Four forms of neuroplasticity
  • Spiking neural nets (so far haven’t had advantages; see zero-divergence inference learning (Oxford) for biologically plausible alternative to backprop)
  • Reinforcement Learning: See Sutton & Barto Chapter 15
  • CNNs seem biologically plausible
  • Brain microcircuits and connectome
  • It seems we need 5-8 layer DNN to approximate cortical neuron as published in “Single cortical neurons as deep artificial neural networks” (Beniaguev et al., Neuron Volume 109 Issue 17, 1 September 2021).
  • Metacognition
    • Sense of uncertainty (potentially unconscious) to be able to doubt oneself
    • Monitoring of actions to recognize when they go off course