Alex Lascelles - Research

Publications

Lahner, B., Dwivedi, K., Iamshchinina, P., Graumann, M., Lascelles, A., Roig, G., Gifford, A.T., Pan, B., Jin, S., Murty, N.A.R., Kay, K., Oliva, A.^†, and Cichy, R.^† (2024). Modeling short visual events through the BOLD moments video fMRI dataset and metadata. Nature Communications, 15(1), 6241. Paper

Gifford, A. T., Lahner, B., Saba-Sadiya, S., Vilas, M. G., Lascelles, A., Oliva, A., Kay, K., Roig, G., and Cichy, R. M. (2023). The Algonauts Project 2023 Challenge: How the Human Brain Makes Sense of Natural Scenes. arXiv, 2301.03198. arXiv Paper ★ Website

Monfort, M., Ramakrishnan, K., Andonian, A., McNamara, B., Lascelles, A., Pan, B., Fan, Q., Gutfreund, D., Feris, R., and Oliva, A. (2021) Multi-Moments in Time: Learning and Interpreting Models for Multi-Action Video Understanding. IEEE Pattern Analysis and Machine Intelligence (PAMI), 44(12), 9434-9445. Paper ★ Website

Cichy, R.M., Dwivedi, K., Lahner, B., Lascelles, A., Iamshchinina, P., Graumann, M., Andonian, A., Murty, N.A.R., Kay, K., Roig, G., and Oliva A. (2021). The Algonauts Project 2021 Challenge: How the Human Brain Makes Sense of a World in Motion. arXiv, 2104.13714. arXiv Paper ★ Website ★ GitHub Code

Ramakrishnan*, K., Monfort*, M., McNamara, B., Lascelles, A., Gutfreund, D., Feris, R., and Oliva, A. (2019). Identifying Interpretable Action Units in Deep Networks. IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2019) Workshop on Explainable AI. Paper

Cichy, R. M., Roig, G., Andonian, A., Dwivedi, K., Lahner, B., Lascelles, A., Mohsenzadeh, Y., Ramakrishnan, K., and Oliva, A. (2019). The Algonauts Project: A Platform For Communication Between the Sciences of Biological and Artificial Intelligence. arXiv, 1905.05675. Paper

Lascelles, A. and Bhattacharya, J. (2018). Neural Correlates of Crossmodal Correspondence Between Pitch and Visual Motion. MSc Thesis. Paper ★ Poster for Behavioral Data

Lascelles, A., Drake, J., and Garraffo, C. (2017). Realistic MHD Modeling of Wind-Driven Processes in Cataclysmic Variable-Like Binaries. MPhys Thesis. Paper ★ Video

Lascelles, A. (2012). Submarine Drag Reduction Study BAE Systems Maritime Summer Internship. Paper

Past Projects

The Algonauts Project 2023

How the Human Brain Makes Sense of Natural Scenes

The 2023 edition of The Algonauts Challenge focused on explaining responses in the human brain as participants perceive complex natural visual scenes. Through collaboration with the Natural Scenes Dataset (NSD) team, this Challenge ran on the largest suitable brain dataset available (brain responses from 8 human participants to in total 73,000 different visual scenes) opening new venues for data-hungry modeling. The challenge was organized in partnership with the Conference on Cognitive Computational Neuroscience (CCN).

At every blink our eyes are flooded by a massive array of photons — and yet, we perceive the visual world as ordered and meaningful. The primary target of the 2023 Challenge was predicting human brain responses to complex natural visual scenes. We posed the question: Given a set of images, how well does your computational model account for the brain activations when a human viewed those images?

Learn more about Algonauts 2023 here.

Example of the online tool we built.

Goal Grounding Project

Group collaboration between teams at MIT and Facebook Reality Labs

The aim of this project was to collect annotations on egocentric (first-person) videos that have a specific goal which is being carried out. This results in visual ground truth data for when a certain goal is taking place within a video. This would be useful for a number of different applications, for example, an AI assistant predicting and helping with the next action/goal a human will encounter.

To do this, our team at MIT collaborated with a team at Meta to build an interface to collect data from participants online, selected suitable data sets to sample from, collected the data using the Prolific platform, and analyzed the data. The interface was a useful online tool that allowed a user to quickly select portions of video while watching, and was flexible enough to be applied to other research aims.

If you'd like to know more about this project please get in touch.

The Algonauts Project 2021

How the Human Brain Makes Sense of a World in Motion

The 2nd installment of The Algonauts Challenge was centered around video perception and event understanding, and was organized in partnership with the Conference on Cognitive Computational Neuroscience (CCN). The challenge focused on explaining responses in the human brain as subjects watched short video clips of everyday events.

For this release, in addition to building the website, I was also involved in the collection of the data set — creating the video data set (1102 3-second videos), learning fMRI techniques, scanning 10+ participants, and making all the data available to the public.

Learn more about Algonauts 2021 here.

The data set consisted of a set of 1000 3-second videos of everyday events, for which we also provided the fMRI brain data associated with humans viewing these videos.

Example of one of the academic paper summaries.

Computational Neuroscience Paper Summaries

I created one-slide summaries of 100+ computational visual neuroscience academic papers, and other resources. It was made so that you could quickly understand the gist of the paper/resources, and easily come back to it later to remember what it was about. They included a short section on the background, info on what the researchers did, what they found, and any other important information.

The Algonauts Project 2019

Explaining the Human Visual Brain: Workshop and Challenge

While at MIT, I helped to build The Algonauts Project — an initiative bringing biological and artificial intelligence researchers together on a common platform to exchange ideas and advance both fields. Inspired by the astronauts' exploration of space, "algonauts" explore human and artificial intelligence with state-of-the-art algorithmic tools. Our first challenge and workshop, "Explaining the Human Visual Brain", focused on building computer vision models that simulate how the brain sees and recognizes objects, a topic that has long fascinated neuroscientists and computer scientists.

Across two competition tracks, fMRI and MEG, we gave participants a set of images consisting of everyday objects and the corresponding brain activity recorded while human subjects viewed those images. Participants competed to devise computational models that best predicted the brain activity of a brand new set of images. The winners were announced at a workshop that we organized at MIT, featuring guest speakers who are some of the leading experts in the field of computational neuroscience of vision, from institutions such as Harvard, UC Berkeley, Columbia, FU Berlin, and DeepMind.

Learn more about The Algonauts Project and the 2019 edition here, on the webpages that I built to explain and distribute the challenges and workshops.

Summary figure explaining the 2019 challenge.

The data set consisted of a set of 92 and 118 images, for which we also provided the fMRI and MEG brain data associated with humans viewing these images.

Topoplots showing brain activity during trials where the spoken English words "up" and "down" did bias participants' perception of motion (congruent) and where they did not (incongruent).

MSc Thesis (2018):

Neural Correlates of Crossmodal Correspondence Between Pitch and Visual Motion

Supervisor: Prof Joydeep Bhattacharya FRSA

Our senses are not independent entities — instead, they work together to build our realities. When one sense influences another, we call this phenomenon crossmodal correspondence. I studied a particular crossmodal correspondence in which ascending/descending pitches as well as the spoken words "up"/"down" (auditory domain) were able to bias our perception of visual motion (visual domain).

I learned MATLAB so I could design and conduct an EEG experiment in which the participant judged the direction of motion of an ambiguous (same upwards and downwards components) moving grating whilst listening to the auditory stimuli. Using behavioural and ERP analysis, I showed that auditory stimuli were enough to bias peoples' perception of visual motion — an ascending tone caused more people to perceive upwards motion, when in fact there was no overall movement in either direction. This was the first time that this effect was studied using brain imaging techniques.

Read my MSc thesis here or view my poster for the behavioural data here.

Speed Reader

Anvil Hack 2018

While at Goldsmiths, I participated in a number of different hackathons including Anvil Hack. During this event, I worked in a team of 3 to make a platform that allows you to convert text into a form that was much more quickly readable. We were inspired by this advert by Honda. Perhaps in the future, devices that use this technology might be commonplace.

An example of this technology in practice. Your reading speed is greatly increased when your eyes don't have to make multiple saccades from word to word, but can instead remain focused in one place.

A close up of one of my simulations.

The magnetic field lines are represented in grey and wind speed by the red/black color map. The blue color map shows mass density (effectively tracing accretion).

MPhys Thesis (2017):

Realistic MHD Modeling of Wind-Driven Processes in Cataclysmic Variable-Like Binaries

Supervisors: Dr Cecilia Garraffo and Dr Jeremy Drake

When main sequence stars such as our Sun orbit a star known as a white dwarf, they can form something called a cataclysmic variable-like binary. The charged ions contained within the magnetic fields of these stars creates a complex magnetic wind structure as the two bodies orbit each other. When this wind structure decouples from the system it carries away mass and angular momentum, which impacts how the stars spin and how their orbits evolve.

Using 3-D magnetohydrodynamic simulations, I explored the effects of orbital separation and magnetic field configuration on these mass and angular momentum loss rates for binary systems. The current models didn't include the magnetic fields of the stars. It was hypothesized that including the magnetic field of the white dwarf in the binary system could alter these rates. I found that when including these magnetic fields, the mass and angular momentum loss rates dropped by factors of 4 and 6 respectively, suggesting that the current models for these systems should be amended to include the magnetic field.

Read my MPhys thesis here or watch a 20-min video of my thesis defense here, given at the Harvard-Smithsonian Center for Astrophysics High Energy Seminar.

Hack @ Brown 2017:

Saguaro: An open source, extensible home automation platform

I attended a hackathon at Brown University where I worked within a team of 5 to design and build an IOT platform called Saguaro. Saguaro allows hardware devices to be controlled by a simple web interface.

Activating hardware in your home — lights, windows, anything you can run a wire to — is as simple as pushing a button. Additionally, Saguaro learns your schedule, and over time will be able to anticipate your actions and automate your home for you.

Learn more about this project here.

A prototype of our IoT home! Built using an Arduino.

Some of the different orbits that I discovered, shown in stationary (left) and rotating (right) frames of reference.

Artist's rendition of an Earth-like planet orbiting a binary star system.

Bottom image credit: Lynette Cook / extrasolar.spaceart.org

Planetary Orbits Around Binary Star Systems

Computing Project #1 (2016)

Incredibly, almost half of the star systems that we see in the sky contain multiple stars! This project aimed to simulate planetary orbits around a binary (two) star system. This is important to study because such simulations could be used to detect habitable planets outside our solar system (exoplanets). For life to exist, the planet on which it occurs must keep a stable orbit over long time-scales. This is similar to the novel "The Three-Body Problem" by Liu Cixen, now made into a Netflix series.

To find stable orbits, I used a Runge-Kutta approach to solve the differential equations involved in such a 3-body system and tested various initial velocities and separations. I found many different p- and s-type orbits (the two species of stable planetary orbit), as well as some chaotic orbits, and discovered whether they possessed a habitable zone where the planet could sustain life.

Read my report on this project here.

The Structure of White Dwarf Stars

Computing Project #2 (2016)

White dwarf (WD) stars are extremely dense objects — the only thing preventing them from collapsing into a black hole is a force called electron degeneracy pressure, which has to do with the fact that electrons cannot be pushed close enough to occupy the same energy state. WDs are very important within astronomy — they appear in many areas of study, including galaxy formation, stellar evolution, and supernovae. Since over 95% of the stars we observe will end their lives as WDs, knowing how they function and how their parameters behave is crucial.

In this project, I created a model for determining the mass-radius variation in WDs, and ultimately found the critical mass of a WD (Chandrasekhar limit), beyond which the electron degeneracy pressure can no longer support the star and it must collapse into a neutron star or a black hole.