Posts

Machine learning for gravitational lensing analysis of massive radio surveys Keywords: Machine learning, gravitational lensing, imaging methods in radio astronomy, artificial intelligence, inverse problems, SKA. Observatoire de la Côte d’Azur invites applications for a PhD position in machine learning for astrophysics. The successful candidate will work with Dr. Simon Prunet and Prof. André Ferrari in J.-L. Lagrange Laboratory, within the ANR funded project TOSCA on weak lensing statistics for cosmology, which investigates synergies between optical and radio surveys, with partners in CEA Saclay, Nice, Caen and Geneva.

Simple Plots.jl receipe for waterfall plots (a la Unknown Pleasures 😀). Install using: pkg> add https://github.com/andferrari/WaterFall.jl and using WaterFall x = [exp(-(n - 1e-2*τ^2)^2/τ) for n in 0:300, τ in 40:10:160] plotfall(x, w=2)

Fourier transforms of a sine wave using a Pluto.jl notebook: The interactive notebook can be run on binder, but be patient 😴. begin using Plots, PlutoUI function DFT_plot(ω₀, N) f(ω) = abs(sin(N*ω/2)/sin(ω/2)) ω_DFT = (0:N-1)*2π/N plot(ω -> f(ω - ω₀), 0, 2π, label="Fourier transform", xlabel="ω", w=2) plot!(ω_DFT, f.(ω_DFT .- ω₀), line = :stem, label="Discrete Fourier Transform", marker = :circle, w=2) vline!([ω₀], label="ω₀ = $(round(ω₀, digits=2))", w=3) title!("Fourier transforms of x(n)=exp(jω₀n), n=1…$N") end end begin ω₀Slider = @bind ω₀ Slider(0:1E-2:2π, default = 0.

First post covid live presentation at SLSIP21!

Martjin Van Ende presentation at 2021 Asilomar conference: Next-Generation Traffic Monitoring with Distributed Acoustic Sensing Arrays and Optimum Array Processing.

A simple noise2self experimennt using a DCNN. The julia code is here.

Martjin Van Ende presentation at 2021 IRIS Marine Seismology Symposium: A self-supervised Deep Learning approach for improving signal coherence in Distributed Acoustic Sensing.

Convergence rate vs Hessian condition number using a Pluto.jl notebook: using Plots, LinearAlgebra, PlutoUI begin f(x) = x[1]^2 + γ*x[2]^2 ∇f(x) = [2*x[1], 2*γ*x[2]] opt_step(x, Δ) = - (x[1]*Δ[1] + γ*x[2]*Δ[2])/(Δ[1]^2 + γ*Δ[2]^2) function grad_desc(; n_itermax = 40, x0 = [10, 1], ϵ = 1e-1) x = zeros(2, n_itermax) x[:,1] = x0 k = 1 while (norm(∇f(x[:,k])) > ϵ) && (k < n_itermax) Δ = - ∇f(x[:,k]) x[:, k+1] = x[:, k] + opt_step(x[:,k], Δ)*Δ k += 1 end x, k end end @bind γ Slider(1:0.

Line by line Deep Learning algorithm for detection of artifacts in uncalibrated multispectral satellite images The aim of artifact detection satellite image is to detect and locate pixels or set of pixels that are decorrelated from the statistically representative background. For instance, in forestry applications, infected trees are considered as artifacts, as well as water pollution in coastal clear waters. Contrarily to the target detection problem where the spectra of the target must be known, artifact detection methods generally assume no prior information about the spectra of the pixels, in both nominal and abnormal cases.

Inspired by the nice tweet of Constance Crozier, here is the julia code to illustrate the fit of a sigmoid as a function of the number of samples. using Plots, LsqFit @. σ(x, θ) = θ[1]/(1 + exp(-θ[2]*x)) x_plot = range(-6,stop=6,length=128) y_plot = σ(x_plot, [1, 1]) n = 32 xdata = range(-6,stop=6,length=n) ydata = σ(xdata, [1 1]) + 0.05*randn(n) anim = @animate for i in 2:n θ_est = coef(curve_fit(σ, xdata[1:i], ydata[1:i], [1.