Mini projects: Please suggest your topics here!

[Florian Marquardt]

Dear all,

If you want to do a mini project to get a certificate, please start by suggesting a topic here. I may still give feedback to revise it (or ask you to look for another one). It is first-come-first-served.

Best regards,

Florian

[Garvit Agarwal]

What is the deadline to give the topic for the mini project?

[rajarsi pal]

Is it possible to choose a topic that is though not in current paradigm of Neural Networks but provides a richer understanding of the dynamics within ?

If so, then I would like to hover over Hopfield Networks, though not very effective for practical purposes it provides a system which could be analyzed within dynamical systems framework, like the fact it has a Lyapunov function that guarantees its convergence, and so on.

I will be glad, if you can provide me a broader view in this regard.

 

[Florian Marquardt]

Deadline would be July 1st, so you have enough time to work on it...

[Arpitha PV]

I would like to work on Convolutional Neural Network in Astrophysics for Identifying Gravitational Lenses.

[Michail Kravchenko]

Hi everyone!

I'm currently working on determining the stars of the signal rising slope for HPGe (High Purity Germanium) detectors using a pre-trained Autoencoder (Conv) + usual Dense part, for particle physics. That's my first ML project, I started to do my first steps in this task just at the beginning of these lections, and there is significant progress since then (actually after lectures on CNN and Autoencoders). If it is possible, I'd like to use this topic as my mini-project. Especially since it's my current work)

Regards,

Michail.

[Florian Marquardt]

Dear Michail, dear everyone,

I forgot to mention: the mini-project should be something different from a project you are already doing anyway. It may be related, but it must not be just a part of that --- otherwise it's like submitting the same project for credit at two different places. Maybe you can find something where you still profit from what you have been doing but it is still its own little stand-alone mini-project.

Best regards,

Florian

[Florian Marquardt]

OK! But since I know there is a paper out there (and maybe even code), please make sure to do it in your own way (even if the results are not as great as the published paper).

[Florian Marquardt]

@rajar... : Hopfield networks are fine, this is an interesting topic. What exactly do you plan to do?

[Leonardo Alvarado]

I would like to apply t-SNA for clustering the synthetic response spectra obtained from numerical modelling of earthquakes in a sedimentary basin.  I would like to use t-SNA for reducing the dimensionality of the response espectra curves and then apply a clustering algorithm (e.g. K-means) for grouping the reduced version of the curves. The goal is to see relationship between geological conditions  of the soil and different groups of the curves obtained .

[Hasan Sansar]

Hi,

I want to work on detection of anomaly in magnetic moment of muon make use of deep neural networks. If this can be done.

7 Haziran 2021 Pazartesi tarihinde saat 20:50:53 UTC+3 itibarıyla jole...@gmail.com şunları yazdı:

[Florian Marquardt]

Dear hasansansar : Could you please describe a bit more what would be the training data (input -> output) that you would want to train the network on, and where you get it from ? (so I can tell you whether I think it will be possible)

[Florian Marquardt]

This sounds interesting. Is this a project you are doing anyway as part of another exam for which you get credit (like some master thesis or similar)? In that case it would be needed that you slightly change this/add something in addition. If you only do it for this course, then it is fine for the certificate just as it is. But you would need to start simple and check for yourself if you, as a human, can see some qualitative difference between the different curves --- otherwise, it will be hard for any technique to extract something meaningful.

[Hasan Sansar]

Dear Florian,

I hope I can use datasets that publishing by Muon g-2 Initiative at Fermilab in April. I think we can get different magnetic moment values from training to training using this datasets.

Best regards,

Hasan

8 Haziran 2021 Salı tarihinde saat 20:18:40 UTC+3 itibarıyla Florian....@physik.uni-erlangen.de şunları yazdı:

[Leonardo Alvarado]

The original project was part of my master thesis. In that opportunity we did the clustering of the Response Spectra (50- component vector) using K-means with eucliadean distance and other algorithm for measuring similarity between curves (Dynamic Time Warping or DTW) for one of the component of the groun motion  (article). In this opprtunity I would like to explore t-SNE algorithm  to reduce the dimensionality before apply clustering algorithm and maybe use the three components of the ground motion .

[Leonardo Alvarado]

Sorry, this teh correct link of the article

[Florian Marquardt]

@joleonar : OK, sounds reasonable!

[Dominik Rattenbacher]

Dear Florian,

I would like write a neural network peakfinder  for the resonances of micro ring and disc resonators. The real data for that is obtained in my PhD work and I have already programmed a working algorithmic peakfinder for that task. The challenge here is that there can be several resonances (different polarizations or splitting of clockwise and counter clockwise modes), and depending on the resonator design the SNR and the width can change by 2 orders of magnitude. Therefore, I just what to see how good a neural network can perform on this and how it depends on its training, e.g. if it is always trained on resonances with one particular linewidth, would it catch one which is 10 times broader or narrower? What happens if it is trained on single peaks, but peaks can get arbitrarily close and start overlapping? etc.
The network would be trained on simulated data and then tested on real experimental data.

Best regards,

Dominik

[ashish arya]

Dear Florian,

     I joined the course rather late. However I want to work on Mini-project as a part of this course. I want to take up the Mini-project on: Exoplanet detection from the Kepler/TESS data using Spiking Neural Network.Exoplanet detection is a major problem in astronomy and I wanted to work on this problem for long. The dataset is available online and Spiking Neural Networks library is available on Pytorch. The challenge will lie in removing false positives. 

    

     What's your opinion?

With regards,

 Ashish K S Arya

On Monday, June 7, 2021 at 4:29:26 PM UTC+5:30 Florian....@physik.uni-erlangen.de wrote:

[stan r]

Hello Florian,

For my project I was hoping to merge the spirit of the auto-encoder challenge with a further investigation into t-sne for clustering datasets. I have some datasets of images that I labelled myself (which I definitely spent too much time on). I was hoping to replicate the technique in the tutorial/homework of training an autoencoder to reproduce them and subsequently take the dataset represented in encoded space and apply a dimensionality reduction technique to try and visualise clusters in the dataset. 

The key point of the investigation is to optimise the encoded representation of the image (in a similar manner to the autoencoder challenge) to see if this methodology could correspond at all with my human labelling. Any patterns analysed beyond my human performance would be a bonus. The structure of the autoencoder and t-sne settings that produce the best results will be a part of my presentation.

I hope to use this project to build skills that will speed up and inform future labelling tasks as well as potentially produce some pretty visualisations that highlight the usefulness of the T-SNE/Dimensionality reduction methods for analysing high dimensional datasets.  

Hopefully this fits the brief.

Thank you for your time,

Stan

[Florian Marquardt]

Dear Ashish, 

interesting suggestion, please go ahead!

Best regards,

Florian

[Florian Marquardt]

Dear Stan,

that is an intriguing idea! I will be curious to see the results! How large is your dataset?

Best regards,

Florian

[Suyash Gaikwad]

Dear Florian,

I want to work on generating potentials in quantum mechanical hamiltonians using NN and using hamiltonian as the loss function and using a random distribution of position and probability densities as input data in an inverse Schrodinger equation approach.

For that, I plan to reproduce results for H2 molecule from the following paper and then extend the analysis using ANNs to predict a double-well potential in ultracold atomic systems, using data sets for images of single-shot measurements, obtained experimentally.

Link: https://arxiv.org/pdf/2006.13297.pdf

The code for the first part is available, but the challenge would be to do the second part.

Please let me know if this works.

Thank you.

Regards,

Suyash Gaikwad

[Shivani Semwal]

Dear Florian,

I plan to use autoencoders combined with the Gaussian mixture model to characterize the colloidal structures in a 2D plane. In this project, I will use my experimental data and train autoencoder to determine the local bond order parameter.

Regards,

Shivani

[Arpitha PV]

What is the deadline for the Mini Project?

[Seowon Choi]

Dear Florian,

Hello, I would like to compare the image classification with using CNN and CNN Quantization for the mini project.

Would this subject suit for the mini project?

Thank you for the help in advance,

Best regards,

Seowon Choi

2021년 6월 7일 월요일 오후 7시 59분 26초 UTC+9에 Florian....@physik.uni-erlangen.de님이 작성:

[Francesco Di Colandrea]

Dear Professor,

in my research activity I am currently focusing on realizing unitary operations by means of a limited set of liquid-crystal waveplates. We will fabricate these waveplates in our laboratories in the next few days. Since I will use the polarization of the light beam as internal degree of freedom, for the time being I am just focusing on two-level unitary evolutions. In these days, with my supervisor we were wondering if there exists some direct optical measurement allowing us to characterize the polarization transformation implemented by the liquid crystals point-by-point on the light beam. We found out that literature is quite rich in this so-called “quantum process tomography”. However, this technique is profoundly based on abstract mathematics (especially group theory and spectral decomposition) and it typically does not determine the explicit form of the 2x2 matrix representation of the process (except very exceptional cases), but an equivalent representation called “Kraus representation” (I’m not very expert on this mathematical issue), which appeared quite useless. We would like to characterize these optical operators in a more “explicit” way, for example determining point-by-point the eigenvalues and the eigenvectors of the process by directly measuring how the incoming light beam changes point-by-point after the liquid crystals. This could be done by preparing different input polarizations (for example horizontal, vertical, diagonal, antidiagonal and circular) and looking at how the intensity has changed point-by-point (i.e. pixel-by-pixel) on the camera after the liquid crystals. By recalling what I learnt from your lectures on training, I was thinking of training a dedicated neural network with a batch containing different sets of numerically generated pictures (for the different polarizations) with the numerically computed eigenvalues and eigenstates (for a great number of non-unknown unitary processes).

Thanks for your attention and for your time, I hope this idea can be fitting for the mini-project, even if I’m not sure it will work properly.

Kind regards,

Francesco Di Colandrea

[Mahnoor Fatima]

Dear Florian,

I would like to work on Statistical Approach to Quantum Phase Estimation using multiple qubits as described in this paper. I want to try to optimize this algorithm using different optimization models and compare their performance.

If this project is not a good fit, can I work with quantum monte carlo simulation?

Regards,

Mahnoor 

[rajarsi pal]

in regard to,  https://groups.google.com/g/machine-learning-for-physicists/c/c6rLV3MllpA/m/DB9zEgFkAQAJ

Hello Prof. Florian,

I have figured two topics and I will be glad to work on either one of them, depending it will suit to the requirements ; 

1. I would like to work on the applications of Hopfield NNs in solving optimization problems. Following from the original work of Hopfield and Tanks it has been shown that such networks may aid in solving combinatorial optimization problems, specifically the TSP(Travelling Salesman Problem) and N-Chess Queens problem, over this, I would like to try and use Hopfield nets for the Stable-Matching problem and compare its efficiency to the standard Gale-Shapley algorithm. Since the Stable-Matching problem is somehow similar to the TSP, I think similar results may follow.

2. It could be on analysing the dynamical behaviour of Hopfield NNs. The original version of Hopfield nets uses symmetric weights between neurons which guarantees its convergence to a stable minimum. However, when this restriction is removed the network often settles on periodic orbits and chaotic attractors. Theoretical analysis on such continuous Hopfield nets (continuous-time evolution) [http://dx.doi.org/10.1063/1.2220476],[Zheng et.al] are available. Over them, I would like to consider the case where the neurons take discrete values and study their dynamics in configuration (presumably Hamming) space, I propose to calculate measures analogous to Lyapunov exponents(by using Hamming distance over euclidean) to quantify its behavior.

I will be glad if you can comment on the feasibility of these ideas and will be moving to work on them if you agree.

[Marta Galbiati]

Dear Professor,

my idea for the mini-project is to use data that I have from particle-in-cell simulations of the laser-plasma interaction. I want to work on a neural network that takes in simulated trajectories of relativistic electrons accelerated during the interaction and try to see if the network can predict laser and plasma parameters  that determine the interaction itself. Since I have  trajectories in input, I will consider using a recurrent neural network.

It would be great if you can comment on my idea.

Thank you for your attention.

Best regards,

Marta Galbiati

[Garvit Agarwal]

For the mini-project, I would like to apply a CNN for classifying galaxies into the broad morphological types of spiral, elliptical, lenticular, and irregular. I am aware that this exact form of work has been done before multiple types, but I will try to think of my own way of doing it. As of now, I am thinking of taking the data of the galaxies of EFIGI catalog from SDSS. Please let me know what you think, thank you. 

[Juan Carlos Badilla Rojas]

Dear professor,

Sorry I've taken so long to suggest a topic. I have been very busy and I had an idea that should have posted earlier but oh well. 

I have previously volunteered with a national non-profit organization that works for the conservation of the Baird's tapir. They use camera traps to study them, and they usually recover the cameras and go through the pictures one by one to see which animals appear and see all the appearances of the tapir. I was thinking of developing a program that learns how to identify which animal appears in the photo, so the organization can use it when they recover a camera. They use the data from the camera traps to write papers and grant proposals, so I'm thinking of maybe including some data analysis aswell. I haven't talked about this with the organization leader but I'm planning to do so if you think it's a good project. I know this has been done before, but I'm not sure if any programs are easily available or if they have learned how to recognize the Baird's tapir correctly. Even though it would have an application outside this course, I'm not getting any course credit or payment from the organization, so I thought this could be a good project to develop for the course and then donate to the organization.

However, I do have a couple of concerns regarding this project. I would really like to know how many photos would be enough for the neural network to be able to recognize the animals (tapirs, felines, small mammals, humans or nothing...), because I'm not sure how many the organization has available. And if they have an amount that is enough to do this project, I'm worried about the time. How much time would we have to develop the project and present it? Because I still need to talk with the organization leader, travel to another part of the country to get a hard drive with the photos, and maybe label all the training samples. 

I would love to work on this because tapirs are my favorite animal and I really support their conservation! However if you don't agree with this topic, let me know. I don't have any other ideas so I would appreciate if you could give us possible projects so I can develop one to get the certificate.

Thank you,

Juan Carlos Badilla

[Shashi Kumar Samdarshi]

Dear Florian Marquardt,

I have discovered a topic for the mini-project form research paper which is Classification of Broad Absorption Line Quasars with a Convolutional Neural Network.

I don't know much what will topic for the project exactly. 

Let me know what you comment on it.

Thanks & Regards,

Shashi Kumar Samdarshi

[Vishal Sv]

Dear Professor,

Sorry for the late reply, I would like to work on classification of volcanoes during the holocene period. I am thinking of using the data from open source dataset "Volcanoes of the world". That dataset has been used for classification by other researchers as well. But I would like to try and do it with my own code and I would like to use the location data to group volcanoes and try to predict future eruptions. I will also look for any other new dataset that could give me more input variables. Please comment your views on this.

Regards,

Vishal S V

[Ali PM Vafa]

Dear Florian,

Regarding the mini-project for this course, I would like to work on using Reservoir Computing architecture for predicting the nonlinear dynamics of ultrashort pulse propagation in optical fibers. For your reference, in this recent paper [1], of which I will get my training and validation data, an RNN architecture has been used for this purpose. I am going to develop software-based reservoirs of different types and compare my results to the results in this paper.

I am looking forward to hearing your comments.

Best

1. Predicting ultrafast nonlinear dynamics in fibre optics with a recurrent neural network

[ayan kumar ghosh]

Dear Professor Marquardt,

Proposal for mini-project: I have chosen passive brain computer interfaces [Zhang et. al]  as my area of mini-project. The idea is to predict the mental workload for a given subject in a particular session based on EEG data obtained from a training data set. The datasets for both the sessions are available in the Link. I am looking forward to hear from you.

Best regards,

Ayan

[Sagar Saha]

Dear Professor Marquardt,

First of all, very sorry for the late reply. I had exams for the last 2 weeks and I was a little busy there.

I'm a beginner in Machine Learning but I'm very much interested. I would do something regarding your suggested topic, last year:

Apply supervised learning to learn something about quantum systems (e.g. predict the underlying quantum state based on observations, or predict the Schrödinger equation solutions based on a given potential or for a small Hamiltonian matrix, e.g. for a two-level system)

I'm going to do some research regarding this and will try to do some mini-project regarding this. But Sir, if you have any specific topic in mind for a beginner like me, please suggest me. Better I would look for that.

Best Regards,

Sagar

[akash kumar singh]

Dear Prof. Florian Marquardt,

I am learning about quantum walks this summer and I have found this really interesting paper. It will be really great to combine and implement what I have learned in this course with this. My proposal for a mini-project will be"Estimating the coin parameters of Quantum walks using Machine learning algorithms from their probability Distributions" 

as described in this paper.  

https://arxiv.org/pdf/2007.04572.pdf

I will try to use different machine learning algorithms to estimate the coin parameters and compare my results with this paper.

Regards

Akash

[Florian Marquardt]

Dear Suyash,

so you plan to do supervised learning where given some quantum ground state or some time evolution you can predict the underlying potential? Sounds good to me --- the interesting part will be your choice of random potentials for the training. I suggest Gaussian random fields, but maybe you have other ideas. This will influence the predictions later.

Best regards,

Florian

[Florian Marquardt]

Dear Arpithav,

the deadline for completion of the mini-project is July 19, because then we will have the presentations, where you present the results in a few slides.

Best regards,

Florian

[Florian Marquardt]

Dear Seowon,

I do not know 'CNN Quantization', but if it is some variant on CNN, you could do it, provided it is not already directly done or at least that you use another data set and implement it yourself.

Best regards,

Florian

[Florian Marquardt]

Dear Francesco,

That sounds interesting: so you want to do supervised training that maps a set of observed pictures (taken for different input polarizations) into what exactly? Do you assume a spatially homogeneous transformation (so in the end you just want to obtain a 2x2 matrix) or do you assume that it can even spatially vary (so you would obtain a whole field) ? Either way, if you think you can properly do the simulation to generate the training data, this sounds like a nice inverse problem to train a NN on.

Best regards,

Florian

[Florian Marquardt]

Dear Mahnoor,

quantum phase estimation paper: I do not quite see what this would have to do with neural networks or any of the techniques we discussed in this lecture. Maybe you can identify a very simple case where a neural network helps to estimate the parameters (such as the phase) inside a unitary by looking at the output states for given input states? So, some unitary U(phi) is applied to a set of state vectors, psi1,psi2,psi3,...,psi_n, and this results in outputs Upsi1,Upsi2,... You could feed these inputs and outputs into the network (as one long vector, splitting real and imaginary parts) and have the network predict for you phi. Maybe to keep it simple, you only want to feed in one input-output pair (psi and Upsi), that should be enough in simple cases to determine the phase phi inside the unitary.

(or anything other where neural networks come into the game).

Best regards,

Florian

[Florian Marquardt]

Dear Rajarsi,

These are both interesting suggestions on Hopfield networks. The Hopfield dynamics could be very interesting, but at the moment I do not know which dynamical law you would set up for the discrete networks (what does it even mean to have a non-symmetric weight in the dynamics of a discrete Hopfield network?). But in my mind, even for the 'simple' model with symmetric coupling and some kind of simple Metropolis-type dynamics, it will be very interesting to see the dynamics, especially if you try to put more and more patterns into the network --- I guess something funny happens then.

Best regards,

Florian

[Florian Marquardt]

Dear Marta,

That sounds like a perfect inverse-problem type of idea! An advanced version would add noise to the trajectories and see whether the network can still come close to predicting the correct parameters.

Best regards,

Florian

[Florian Marquardt]

Dear Garvita Kagarvit,

Maybe you could change it a bit by trying to train an autoencoder to reproduce the galaxy pictures? If there are sufficiently many of these pictures, this could work, and it would be really interesting to see how well it does (and maybe even what the latent space variables mean). You could start from the kind of architecture we had considered in the tutorials.

If you prefer the classification problem, then I would suggest you add some other variables that have been measured, if there are any, apart from the human-made classification into types.

Best regards,

Florian

[Florian Marquardt]

Dear Juan,

That sounds very interesting.

Since the presentations are already on July 19, it will probably be not sufficient to do all of this.

But I am sure you can find on the web some kind of image database with pictures of different animals (or even in the ImageNet and similar databases there are many animal pictures, maybe one can download them individually). Then you could train on such a database as a test case and find out how many pictures you need during training to get a good distinction (this will depend on how many categories you want to distinguish). This you can present as a mini-project and then later replace the pictures with those that you really will use for this particular task.

Best regards,

Florian

[Florian Marquardt]

Dear Sashi,

If you can get access to the raw data (from the database they used) then you could indeed train a neural network. Keep it simple, only do some classification task or simple regression (predicting numerical values of parameters), don't use everything in there. Do not use any of the more advanced ideas they have in the paper.

Best regards,

Florian

[Florian Marquardt]

Dear Vishal,

That sounds like an interesting dataset. The question is what would be your input and what would be the desired prediction. I was not able to find the actual data set. But you have to make sure you select the input data so there is at least a chance of making a good prediction. (If the input is just a single number, and the output also a single number, it would be a first step to plot one vs the other and see whether there is any rough correlation)

Best regards,

Florian

[Florian Marquardt]

Dear Ali,

Thank you - that sounds like a good project, but please keep it simple. So I guess the idea in reservoir computing is to take a random complicated transformation of the input and then put a simple network on top and do supervised training. What would be the input and what the output? Is the input the previous time series and the output some prediction about the future evolution?

Best regards,

Florian

[Florian Marquardt]

Dear Sagar,

How about this:

Produce a random real symmetric NxN matrix (N not too large, maybe order 3 to 10), e.g. by taking every entry as a Gaussian random variable and symmetrizing it, and consider that as a Hamiltonian H. Calculate the Schrödinger evolution of a wave function, according to i dPsi/dt = H Psi, with Psi as a complex N-dimensional vector, for a short time interval T, starting from an initial state Psi(t=0) that is always the same, namely (1,0,0,0,0,..). Now train the network to predict Psi(T) given the Hamiltonian, where you can feed all the entries of the Hamiltonian matrix as input to the network. Alternatively, try to solve the inverse problem, where you take as input Psi(tau), Psi(2*tau), Psi(3*tau), ... (with at least N time points) and the output should be the Hamiltonian. As usual, complex quantities should be split into real and imaginary parts. For the time evolution scipy has routines, or you can use the RungeKutta we used in one of the tutorial examples.

Best regards,

Florian

[Florian Marquardt]

Dear Akash,

That sounds interesting. Please concentrate on one algorithm (neural networks) and try to generalize at least slightly. For example, you can take a different unitary at each position, and try to estimate all these parameters simultaneously.

Best regards,

Florian

[Garvit Agarwal]

Dear professor,

Using an autoencoder on the galaxy images sounds like a cool idea to me. I would like to explore this instead of the classification problem. What did you mean by latent space variables? I have about 1,40,000 images, having all kinds of galaxies and possibly many not-so-ideal images but this should be a big enough dataset. What do you think?

Best regards,

Garvit Agarwal

[Ali PM Vafa]

Dear Florian,

Thank you for your response. Regarding the concept of reservoir computing, yes the basic idea is to use a fixed random nonlinear transformation of the input to a high-dimensional space, on the condition that the transformation is complex enough, and then use a simple FFN on top that is adjusted by supervised learning. In the case of this particular project, the input is the temporal intensity profiles at 10 consecutive cross sections of the fiber separated by a sampling distance dz and the output is the intensity profile at the next cross section. The training data has been obtained by numerical simulation and is available on Gitlab.

Best,

Ali

[Florian Marquardt]

The latent space variables are the neurons in the bottleneck.