Strong Lens Objects

Title: Looking for Strong Lens Objects in the Sky [Archived Project]

September 2016 - December 2017


Within Cosmology, strong lensing phenomena refers to an interaction between a light source visible to an observer, such as a star, and a highly massive object with strong gravity positioned between the object and the light source, usually a galaxy. The presence of the massive object warps the path of light travelling from the source to the observer, often distorting the original image of the source. From the perspective of the observer, such as Earth, this distortion can take the form of multiple images, bright arcs, and luminous rings. Thus, a “strong lens” visible to the naked eye often appears as a bright central object surrounded by these various distortions, creating beautiful visual effects.

Besides being stimulating to the eyes, these phenomena are scientifically interesting to physicists, because they provide useful “probes” for measuring the distribution of dark matter in the universe. The shape of a strong lens is determined only by a small number of factors, including the distribution of matter and dark matter in its locality. This allows cosmologists to estimate the structure of the surroundings of the strong lens using fewer approximations and assumptions.

As useful as strong lenses are, they are difficult to employ due to their rarity and the absence of an efficient and reliable vetting method. It is estimated that less than 0.001% of all heavenly bodies are strong lenses, and conventional techniques for filtering them from other objects often target very specific types of strong lenses or exhibit competency from one perspective at the cost of another (for example capturing nearly all the strong lenses in a test sample at the cost of an extremely large number of non-lenses in the objects selected by the algorithm).

The latest approach to automated detection of strong lenses comes from machine learning – namely convolutional neural networks (CNNs). CNNs have been shown previously to be highly competent at tasks requiring the algorithm to classify an image as belonging to one of two classes, positive or negative in some way. Some current cosmology literature revolves around specializing CNNs towards classifying images as either containing a strong lens or not containing one, an idea which we contribute to. Our work investigates the effects of preprocessing and CNN optimization techniques on strong lens detection from a cosmological perspective. We conduct experiments using well-known machine learning methods such as regularization, batch normalization, and convolutional filters, and quantify our success not only in terms of machine learning performance metrics, but also through intuitive representations of the experiments, and the cosmological parameters of strong lenses such as Einstein radii, light intensity, and flux lensed image.

Leader(s): Daniel Zhang (slack: @danielz, email:

MDST Participants: Arya Farahi (@aryaf), Pradyumna Gadepally (@pgad), Samuel Tenka (@samtenka)

External Collaborators: Brian Nord (@nord - Fermilab)

Link to the Code:

Link to the Dataset: