Knowledge Vault 2/13 - ICLR 2014-2023
Terrence Sejnowski ICLR 2015 - Keynote - Beyond Representation Learning
<Resume Image >

Concept Graph & Resume using Claude 3 Opus | Chat GPT4 | Gemini Adv | Llama 3:

graph LR classDef embedding fill:#f9d4d4, font-weight:bold, font-size:14px; classDef cortex fill:#d4f9d4, font-weight:bold, font-size:14px; classDef retina fill:#d4d4f9, font-weight:bold, font-size:14px; classDef images fill:#f9f9d4, font-weight:bold, font-size:14px; classDef learning fill:#f9d4f9, font-weight:bold, font-size:14px; classDef basal fill:#d4f9f9, font-weight:bold, font-size:14px; classDef evolution fill:#f9d4f9, font-weight:bold, font-size:14px; A[Terrence Sejnowski
ICLR 2015] --> B[Embedding networks in levels,
operating system. 1] A --> C[Evolution explains biology. 2] C --> D[6-layer cortex enabled expansion. 2] A --> E[24 ganglion types encode
light changes efficiently. 3] A --> F[Neuromorphic camera: waveforms to
spikes, low-power processing. 4] A --> G[Natural images: scale invariance,
1/f^2 power spectrum. 5] G --> H[Percolation reveals phase transition
in image information. 6] G --> I[Renormalization analogous to deep
learning, critical point. 7] A --> J[Priming: powerful one-trial,
unconscious, long-lasting learning. 8] J --> K[Repetition suppression in IT
sharpens representations. 9] A --> L[Homeostatic plasticity maintains firing
rates, enables sparse coding. 10] L --> M[Hebbian + homeostatic plasticity
extracts features via ICA. 11] A --> N[Temporal difference learning updates
weights by reward prediction. 12] A --> O[Basal ganglia: convergent inputs,
dopamine outputs to cortex. 13] O --> P[Three basal ganglia loops:
sensory, outcomes, plans. 14] A --> Q[Brain uses many simultaneous
learning algorithms. 15] A --> R[Synaptic strength: 5 bits
precision per synapse. 16] R --> S[Low release probability synapses
implement dropout regularization. 17] A --> T[Evolution optimized brain's learning
and memory solutions. 18] A --> U[Hippocampus instantiates past, reward
systems plan future. 19] A --> V[Unconscious processing like priming
reveals cortical learning. 20] A --> W['Gut responses' depend on
experience, complement reasoning. 21] A --> X[Visual system encodes more
than just visual stimuli. 22] A --> Y[Superior colliculus: convergent inputs
orient animal. 23] A --> Z[Cortex projects to
striatum's parallel pathways. 24] A --> AA[Nature evolved synergistic forms
of synaptic plasticity. 25] A --> AB[Visual system builds more
than sensory world model. 26] G --> AC[Natural images' percolation has
0.6 critical exponent. 27] J --> AD[Priming strengthens high, weakens
moderate activity connections. 28] O --> AE[Basal ganglia: actor-critic
reinforcement learning architecture. 29] A --> AF[Brain enables fast unconscious
and slow complex responses. 30] class A,B embedding; class C,D,J,K,Q,U,V,W,AA,AD,AF learning; class E retina; class F,G,H,I,AC,R,S images; class L,M,T,X,AB evolution; class N,O,P,Y,Z,AE basal;


1.-The lecture has two themes: embedding networks in a levels diagram and embedding networks into an operating system.

2.-Nothing in biology makes sense except in light of evolution. The 6-layer mammalian cortex enabled cortex expansion.

3.-The 24 ganglion cell types in the retina provide input on light intensity changes, enabling efficient representation.

4.-A neuromorphic camera encodes continuous waveforms as spike sequences, enabling velocity computation and efficient, low-power processing.

5.-Natural images have scale invariance and a 1/f^2 power spectrum, reflecting power law scaling.

6.-Percolation theory applied to natural images reveals a phase transition in information content around the middle bit planes.

7.-Renormalization group theory has a formal analogy to deep learning architectures, suggesting a critical point may be involved.

8.-Priming is powerful one-trial, unconscious learning that can last for years, reorganizing cortical representations.

9.-Repetition suppression in IT cortex after priming may sharpen representations by increasing some connections and decreasing others.

10.-Homeostatic plasticity scales synapses to maintain neuron firing rates, enabling efficient sparse coding of features.

11.-Combining Hebbian and homeostatic plasticity enables nonlinear ICA to extract features like horizontal and vertical bars.

12.-Temporal difference learning uses dopamine to update weights based on reward prediction errors for reinforcement learning.

13.-The basal ganglia receives convergent cortical inputs and outputs to dopamine neurons that innervate the entire cortex.

14.-Three cortico-basal ganglia-thalamic loops associate sensory input with effectors, outcomes, and plans on different timescales.

15.-Many distinct learning algorithms like priming, homeostatic plasticity, and temporal differences occur simultaneously in the brain.

16.-The volume of synaptic spines is proportional to synaptic strength, with ~5 bits of precision per synapse.

17.-Most synapses have low probability of release despite high strength precision, possibly implementing a form of dropout regularization.

18.-Evolution has discovered optimal solutions to many problems in learning and memory that we are just beginning to understand.

19.-Past experience, instantiated by the hippocampus, is evaluated by reward systems to plan future behavior.

20.-Processing inaccessible to consciousness, like priming, reveals extensive unconscious learning in the cortex.

21.-Unconscious, fast "gut responses" depend on experience and are optimized by evolution for survival, complementing slower explicit reasoning.

22.-Analyzing the visual system reveals it encodes more than just visual stimuli, including value information.

23.-The superior colliculus receives convergent cortical inputs and helps orient the animal based on sensory and internal inputs.

24.-The entire cortical mantle projects to the basal ganglia striatum, which has parallel pathways supporting different functions.

25.-Nature has evolved dozens of forms of synaptic plasticity that work together synergistically, beyond just backpropagation or Boltzmann learning.

26.-Predictive priming experiments show the visual system does more than just build a model of the world from sensory input.

27.-Percolation occurs in natural images with a critical exponent of 0.6, suggesting a phase transition in image statistics.

28.-A sharpening effect occurs after priming where strongly active neurons increase connectivity while moderately active neurons decrease connectivity.

29.-The basal ganglia compute state value used for updating weights and making decisions in an actor-critic reinforcement learning architecture.

30.-Brain architecture enables fast, unconscious responding for experts via the basal ganglia vs slow, explicit reasoning for complex decisions.

Knowledge Vault built byDavid Vivancos 2024