Matthew Scourfield
Gather.town id
MLA03
Poster Title
Maximum information retrieval from DESI spectra
Institution
University College London
Abstract (short summary)
The DESI survey will collect data on a large number of galaxies. We shall present our work on techniques to maximise the physical data retrieval for the purpose of galaxy evolution analysis. We focus on the use of machine learning algorithms as a means of intelligent noise reduction, primarily auto encoder methods. In addition, we make use of the latent space representations produced by such methods to automate the selection of similar spectra for stacking.
In anticipation of the DESI SV data we first train our models using SDSS spectra with high signal to noise, adding artificial noise to create a training set. Once the SV data are available we shall look into both training the model on the data, and also at transfer learning between the two datasets.
In anticipation of the DESI SV data we first train our models using SDSS spectra with high signal to noise, adding artificial noise to create a training set. Once the SV data are available we shall look into both training the model on the data, and also at transfer learning between the two datasets.
Plain text (extended) Summary
Maximum Information Retrieval From Optical Spectra
Matthew Scourfield
matt.scourfield.18@ucl.ac.uk
Slide 1:
This project investigates improving the SNR of galaxy spectra collected in the optical
surveys, such as DESI, using variational autoencoders (VAEs), a type of neural network.
Autoencoders work by using an encoder to reduce a multidimensional input down to some smaller number of dimensions.
A decoder then takes these latent dimensions and uses them to reconstruct the input, or in this case a de-noised version of it.
We use 8,000 SDSS spectra to train our model and a further 2,000 for testing. These are
processed by adding artificial noise as well as de-redshifting and normalising them.
Two models are produced, a convolutional and a dense model, each of which uses two layers in the encoder and decoder. The output layer has a sigmoid activation function, all other layers use ReLU functions.
Figure 1a:
This figure shows the general architecture used by out model: An input layer, an encoder made of two layers, a layer for the latent space, a decoder made of two layers and the output layer.
Slide 2:
The convolutional model uses two convolutional layers in the encoder, and two transposed convolutional layers in the decoder.
This model performs better at de-noising, producing spectra with continuum close to the general shapes of the SDSS spectra, though with less noise overlayed.
The model is also capable of reliably identifying the position of several spectral lines and in the case of those such as H⍺ even distinguish between cases of absorption and emission; however, the exact amplitude of the reproduced lines is not always accurate.
Figure 2a:
A comparison between the noisy input, SDSS and VAE reconstructed spectra are shown for 5 galaxies. The reconstructed spectra match the continuum of the SDSS spectra though appear smoother.
Figure 2b:
This figure shows the same spectra as in 2a, but zoomed in to a window around the H alpha line. The positions of the line are well reproduced, though exact amplitudes differ.
Slide 3:
The dense model uses a pair of dense layers in both the encoder and the decoder. A dropout rate of 20% is also used with these layers to prevent overfitting.
The latent space produced by the dense model more effectively
separates out the different classifications of SDSS spectra in the data set, despite these not being used to train the network.
Narrow-line AGN are not separated out, due to their low numbers and similar spectra to emission-line galaxies.
Figure 3a:
A corner plot of the latent space representations of the spectra. Four subclasses of spectra are included based off the SDSS DR14 SpecObj catalogue, quiescent galaxies (9435), emission-line galaxies (5260), narrow-line AGN (656) and broad-line AGN (229). In the various projections it can be seen that the subclasses are separated in the space, other than the narrow-line AGN which overlap with the emission-line galaxies.
Slide 4:
Using the latent space, we can identify similar spectra by looking at their proximity within the space.
This can be used to create automated stacking methods, without the need to manually select spectra.
VAE reconstructed spectra tend to have less continuum noise, while stacked spectra more accurately reproduce line amplitudes.
Figure 4a:
The first plot compares the noisy spectra to the VAE reconstructed spectra. The second plot compares the noisy spectra to the stack produced using the 10 nearest in latent space. The final plot compares the reconstruction and the stacked spectra to the SDSS spectra. The general shapes of the two methods are similar, with the reconstruction continuum being smoother, while the stacked spectra matches SDSS fluxes more closely.
Matthew Scourfield
matt.scourfield.18@ucl.ac.uk
Slide 1:
This project investigates improving the SNR of galaxy spectra collected in the optical
surveys, such as DESI, using variational autoencoders (VAEs), a type of neural network.
Autoencoders work by using an encoder to reduce a multidimensional input down to some smaller number of dimensions.
A decoder then takes these latent dimensions and uses them to reconstruct the input, or in this case a de-noised version of it.
We use 8,000 SDSS spectra to train our model and a further 2,000 for testing. These are
processed by adding artificial noise as well as de-redshifting and normalising them.
Two models are produced, a convolutional and a dense model, each of which uses two layers in the encoder and decoder. The output layer has a sigmoid activation function, all other layers use ReLU functions.
Figure 1a:
This figure shows the general architecture used by out model: An input layer, an encoder made of two layers, a layer for the latent space, a decoder made of two layers and the output layer.
Slide 2:
The convolutional model uses two convolutional layers in the encoder, and two transposed convolutional layers in the decoder.
This model performs better at de-noising, producing spectra with continuum close to the general shapes of the SDSS spectra, though with less noise overlayed.
The model is also capable of reliably identifying the position of several spectral lines and in the case of those such as H⍺ even distinguish between cases of absorption and emission; however, the exact amplitude of the reproduced lines is not always accurate.
Figure 2a:
A comparison between the noisy input, SDSS and VAE reconstructed spectra are shown for 5 galaxies. The reconstructed spectra match the continuum of the SDSS spectra though appear smoother.
Figure 2b:
This figure shows the same spectra as in 2a, but zoomed in to a window around the H alpha line. The positions of the line are well reproduced, though exact amplitudes differ.
Slide 3:
The dense model uses a pair of dense layers in both the encoder and the decoder. A dropout rate of 20% is also used with these layers to prevent overfitting.
The latent space produced by the dense model more effectively
separates out the different classifications of SDSS spectra in the data set, despite these not being used to train the network.
Narrow-line AGN are not separated out, due to their low numbers and similar spectra to emission-line galaxies.
Figure 3a:
A corner plot of the latent space representations of the spectra. Four subclasses of spectra are included based off the SDSS DR14 SpecObj catalogue, quiescent galaxies (9435), emission-line galaxies (5260), narrow-line AGN (656) and broad-line AGN (229). In the various projections it can be seen that the subclasses are separated in the space, other than the narrow-line AGN which overlap with the emission-line galaxies.
Slide 4:
Using the latent space, we can identify similar spectra by looking at their proximity within the space.
This can be used to create automated stacking methods, without the need to manually select spectra.
VAE reconstructed spectra tend to have less continuum noise, while stacked spectra more accurately reproduce line amplitudes.
Figure 4a:
The first plot compares the noisy spectra to the VAE reconstructed spectra. The second plot compares the noisy spectra to the stack produced using the 10 nearest in latent space. The final plot compares the reconstruction and the stacked spectra to the SDSS spectra. The general shapes of the two methods are similar, with the reconstruction continuum being smoother, while the stacked spectra matches SDSS fluxes more closely.
URL
matt.scourfield.18@ucl.ac.uk
Poster file