tc.txt

Experiment at https://127.0.0.1:8000/dashboard

Create FPGA-based emulator (verilog?) if necessary

functions:
    control body
    interactions (social behavior)

can we apply research done on humans to these
    -timeless neuron firing
    -neurotrasmitters
    -neuronal development

information theory for deep learning <-> schmidhuber
" " ...                         " <-> testing in c elegans?
can we then find an 'optimal' network architecture?
Not morphic but graphic (not wire cost economics)

invertible functions in dl - experiment
make algorithm that 'looks' at c. elegans' behavior and reproduces the network weights


Questions:
    NCP inspired by C. Elegans
        Are they robust to adversarial examples (why or why not)
        Are they generalizable to larger networks
        What time overhead do we have because of the nonhomogeneity
        Is the architecture robust to changes (smooth adaptability space)
        Can we have competing results by some certain type of homogeneity of components
        Do we understand the implications of small networks (capacity, speed, etc)

        How does it relate to spike neurons in nets (related to binarized nets)

    Biological
        Can you play the keys at the same time? -? https://www.researchgate.net/publication/258058406_Impaired_Hippocampal_Ripple-Associated_Replay_in_a_Mouse_Model_of_Schizophrenia

        What is the level of abstraction do we need (in the model) for decent behavior controlling (apart from the phenomic dispositions) What classical/interesting experiments, e.g. visual similarity matrix done on mammals/humans and ai systems (haha no visual system), statistical measures of nets (after pruning?), etc, can be transferred to cele.
        Can we incorporate a visual system, or are the variables receiving (how) somehow the
        correct or sufficient representation of its environment for its goals?
        
Read:
    on dp
    zero shot learning - new hot thing, openai/deepmind
    Hierarchical rl
    compressed sensing
    sergey levine papers
    temporal difference models
    NAF


--
off-policy does not converge
thompson sampling
exploration bonus
schotastic energy based policies

ddpg
trpo
SAC
self supervised
supervised meta learning
temporal convolution
are cnns invariant to face location?
paper thing symbolic systems (good things)
btw, that somehow is still in my brain (do more things like that?)
compare those systems
what tools to we have to compare algorithmic approaches in AI

would a fuzzy boundary (what level) and a probabilistic image segmentation (what level) task be the same? (at what level)

svms
approximations solomonoff's theory
polynomial find given points algorithm
more papers on dp folder
small world networks https://pubmed.ncbi.nlm.nih.gov/17079517/
reptile (MAML)
mathematical grounding for anns (what does that mean?)
https://arxiv.org/pdf/2009.10713.pdf

experimentation reflection - https://www.cell.com/cancer-cell/pdf/S1535-6108(02)00133-2.pdf

very nice blog - https://lilianweng.github.io