智力活动是一种生活态度
2023 年 3 月 12 日
.tex.phy
some physical learning that does not have to be in equilibrium
intro to differential geometry: Kamien, Rev Mod Phys 74:953 (2002)
WeiKang Wang, in silico vector field analysis
Waddington landscape 1957
Joanathan Weissman Lab (WI, MIT)
Boston U, Pankaj Mahta Group.
extension of work from Lang et al PLoS comput bio (2014)
found an order parameter, better than UMAP
miimally frustrated topology
proten and mRNA noise levels are Boltzmann distributed.
spin glass: frustration
free energy
gene expression noise can store information
cell types in multicellular organisms
master regulator as order parameter
Hopfield model
forward and backward conditional probabilities
tissue homeostasis: birth and death
cellular Potts model
there is Hamiltonian!
CompuCell3D software
cell size distribution plot is very good
UWashington
collision during cell cycle, replication and transcription
F-mean
bacterial persistence
CRISPER-dCas12a interference
massive parallel assay
Reversibiliy and cell division dynamics of elongated Escherichia coli cells obtained at high pressure
a new cell division model
GCDM
continuous time Markov chain model?
microbial colonizing and phenotypes
Danio rerio
virio Z20
IPTG indeces lac operon
$f(t)=1-e^{-\frac{ln(2)}{t_{1/2}}(t-t_0)}$
cell competition
homeostasis pressure
Adilson E. Motter, Northwest University
contagion
whack-a-mole effect
mono-layer networks: a network of ownership, a network of credit contracts between financial institutes
multi-layer networks:
$\dot r = \alpha r(1-\frac{r^2}{R^2})$
$\dot \theta = \omega$
“decoder”
Leypunskiy 2017, Pittayakanchit 2018
slime molds sense concentrations of fructose in environments
hypotheses:
consumer-resource model: species abundance is solely dependent on resource functuations
Taylor’s law: distribution of abundance change is fit by a power law.
“their model” coarse grained consumer-resource. can predict many kinds of , different resource competition regimes.
Alexandra Horowitz.
Social play
finished soon after I arrived
vector institute
BraXXXX, VQE
Finite time thermodynamics: compete between efficiency and power
energy change between two harmonic oscillators
differential programming, RL-like scheme
missed the flawed thermodynamic definition
reservoir computer: a kind of neural network, for forecasting dynamical systems but most of the parameters are chosen randomly. cheap but works well.
ESN Jaeger 2001
classic representation theorem by WOLD theorem
works because time soaks randomness. equivalent to logical NVAR (non-linear vector autoregression)
enjoys universal approximation theorem, even linear with non-linear readout.
VAR: vector autoregression
VMA: vector moving average
DMD
macket-glass equation as its example
also reservoir computer.
silicon computers represent with binary and compute sequentially.
neural computers use continuous symbols and compute distributedly
dRAM: dynamical random access memory
store a reservoir into another reservoir
another hot topic is Koopman’s theory
RhINNs (PhysicsINNs): inform the NN with underlying physics/rheology
his model is called RhIGNet, and an improved “multi-fidelity RHIGNets”
same group as previous one.
model works better as it goes more complex, while data wants model to be simpler.
we should prioritize data range over size
robot cars (RRP) climbing sand slopes
hopper flow
model manifold: for a given data sampling method, points in parameter space can be mapped to a sampling space
manifold boundary approximation method
supremum principle
Wnt signalling network
residual networks’ has Griffith phase, due to the design (sth. and renormalization)
math heavy
Doi representation, Jarzynski relation, Doi-Peliti field theory
generalized fluctuation-dissipition into non-equilibrium
no coarse graining or slow modes
$\Delta S_M = \frac{Q}{k_B T}n = \gamma n$
also math heavy
detailed TFT: symmetry of particle in vortex field is protected by topology
TUR: thermodynamic uncertainty relation
$\frac{var(j)}j{}\ge \frac{2}{<\Sigma>}$
probability distribution → rate function. don’t know how
“caveat” seems an abused word, another is “ansatz”
detailed balance implies a reversal symmetry
expressed as Markov process: $\frac{p_i}{p_j} = \frac{l(i,j)}{l(j,i)}$
his work is just cycles in non-equilibrium systems
fluctuation theorem for cycle counts. PRE 2021
foundation of thermodynamics, statistical mechanics, can be further founded upon classical or quantum mechanics
three times of Legendre transforms of internal energy ought to give 0, which breaks the reversibility of Legendre transform.
T. Hill 1963: added some sub linear term
they used mean of finite measurements
their understanding of thermodynamics is “nothing more than theories of probability”
entropy production rate
in 1 dimension Derrida)
dimension matters and data reduction may have non-trivial effects.
phase transitions
phase transition again
Mpemba effect: hotter water freeze faster than cold water
totally absent-minded
electrolytes, concentrated solution theory: Gibbs-Duhem relations
very math heavy, and words too small on slides. presentation visualization not so good.
extended DFN model
Onager’s regression
non-linear reaction: at least second order reaction
reactions → emergent cycles
state space renormalization group
second quantization, field operators in multiparticle Fock space
classic particles in quantum language
the operation is the Wick contraction
| $\dot I = D_{KL}[P(x\rightarrow x’) | P(x’\rightarrow x)]$ |
“entropy production” again: Prigogine 1947, Shcnakenberg 1976, Skinner Dunkel 2021
observe k elements in the system: interaction irreversibility $\dot I^{(k)}_{int}=\dot I^{(k)} - \dot I^{(k-1)}$
noise can ruin cellular info sensing
individual cell senses better than population(?) conditional mutual information reflects this
energy-limited vs. nutrient limited
law of energy in biology; model organism being anaerobes B. theta
PPi (pyrophorsphate) as energy source
firefly, one research about vocabulary, one about spatiotemporal pattern
bunblebees, SLEAP to identity action and track
ant, fire ant, BOBbots, multi-occupancy lattice gas, rule based model→observed result
social polarization: vector force, but how agents are embedded into this space? no answer
HKB system, third party induced bistability, symmetry breaking “HOW”!
Team formation: Modeling the Catalysis of Collaboration at In-Person and Virtual Conferences: Non-linear memory model, scialog dataset
mirror game: no designated leader experts reach higher accuracy and velocities
Granger causality analysis: not so advertised. experiment with flow switch. Presentation not so well given.
Bacteria reshape their surroundings to enable migration: use bioprinting. different pore size in media. non-motile vs. motile. motile E.coli can escape tight pores
scalar.seas.upenn.edu
large scale behavior (trajectory): variation explained by exploration-exploitation
fine scale (posture): decompose into 5 primitive postures
connection:
“hopping dynamics” again
non-ergodic drives that prevent the dynamics from relaxing to the steady state within measurement time
the landscape the worms seeking food is also time dependent
this group’s interested is very scattered.
first part just missed: learn macroscopic params in hydrodynamics as fields
Unet to predict some cell protein density, identify relevant signals
model: Oakes et al, BioPhys J 2014
fruit fly tissue
physics-informed machine learning again
Unet also here. turbulent-flow net. basically a new network structure based on unet.
group equivariance leads to another net called Scale Equ-ResNet in order to retrieve symmetry
spatial-temporal neural p…: Baysian Active Learning
1K sensitivity of a channel; 1mK sensitivity of neurons
independent channels, over time of several independent measurements
hypothesis: TRPA1 channels are embedded into a dynamical system near bifurcation. activated channels activates other channels
question: does coupling break the independence of measurements? single channel info is ~1
expanding epithelium,when the area grows linearly, the size only ~$\sqrt{a}$, does this feedback to cells?
cells divide at a smaller size
lower bound of cellular size might come from genome size. CCND1labels DNA demage
no mechanisms yet
a lot of models
key features not heard clearly. adaptive spiking
the way they defined topology number = -1 is not as my expectation
Francesca Serra group experiment
not too much work
Bo Sun again. granger causality again
used a famous model not written down. Sounded like Nagumo model
starfish embryos
gut motions and nerves in crafish
just using 5-HT is not enough to compensate the cutting of nerve
Cramer-Rao bound by maximum likelihood estimation
benchmark was called tug-of-war model
contact inhibition of locomotion
E.coli chemosensing: high cooperativity, perfect adaptation, large noise
symmetric spreading of CheA
MDCK monolayer
very biology heavy
a low fluctuation phase
impedance = pressure/height, WKB approximation
hair cells are the sensors, undergo Hopf bifurcation
mouse brain hippocampus, 2000 neurons
in subgroups, swap 1/2 of the cells to minimize pairwise correlation coefficient
Entropy generation during computation: Szilard’s argument
erase and copy operation of ribonsome works like a universal turing machine.
in theory, logically irreversible process can be thermodynamically reversible, like erasure
MIT media lab
| $\sigma(\vec x_f)\leftrightarrow DK(P(\vec x_f) | P(\vec x))$ KL Divergence |
2 models, 3 by 3 nodes with some lines connecting some pairs, didn’ get what they mean.
they call the thing NESS (non equilibrium steady state).
takehome: dissipative not monotuned increasing as info transmission increases. some channel more efficient than others
diameter variable particles as intelligent agents
diameter function $D(\vec p)$ symmetry determines symmetry of new distribution
policy optimization, reinforce learning to search behavior space
smart vs. inert: ability to perform temporal pattern recognition
a polymer that has multiple foldable positions.
dual scale master equation
information: $lim_{T\rightarrow \infin}\frac{1}{T}\int MI(I,O)dt$
diffusion of signal,some equations, etc, parameters include size of source and receiver
implications:
Landauer principle
D. Woods. Nature 2019
diamond shape 2-in2-out tiles that
Langevin equation
NESS again
Liouville says Halmitonian system must be incompressible in phase space
$H_{ers} = H_0+gH_{contraction}$
micro-canonical energy shell
not following, deterministic finite automata
wanted to follow but failed
also used KL divergence here.
looks like there has been no experiment, maybe I missed it.
Stochastic thermodynamics: ST
fractional Brownian motion: fractional gaussian noise (no idea what)
fluctuation-dissipation is broken by memory in noise
DNA replication error rate 10^-8 ~10^-10
experiment K_D/K_C = 10^-2, observed error rate 10^-3~10^-4, discrepancy
Pareto optimal fronts of kinetic proofreading
generalized Hopfield model
speed-dissipation trade off again
Size→nutrition gradient→differentiation
Multicellular yeast! T yeast
Bozdag
Non-reformable→mechanical challenge
The aspect ratio increases significantly after the size grows
Aggregation not happening; Tangling spans the bulk