Analytical Connectionism

School on Analytical Connectionism — 2023 —

Gatsby Computational Neuroscience Unit · London 28 Aug – 7 Sep 2023
Gatsby Computational Neuroscience Unit
Fig. 1.Gatsby Computational Neuroscience Unit, host venue, 28 Aug – 7 Sep 2023.
Dates
28 Aug – 7 Sep2023
Speakers
10 lecturers
Location
Gatsby Computational Neuroscience Unit
1

Overview

Analytical Connectionism is a 2-week summer course on analytical tools, including methods from statistical physics and probability theory, for probing neural network models of higher-level cognition. The course brings together neuroscience, psychology and machine-learning communities, and introduces attendees to analytical methods for neural network analysis and connectionist theories of higher-level cognition and psychology.

Connectionism, a theoretical approach in cognitive science, uses neural network models to simulate a wide range of phenomena, including perception, memory, decision-making, language, and cognitive control. However, most connectionist models remain, to a certain extent, black boxes, and we lack a mathematical understanding of their behaviours. Recent progress in machine learning theory has provided novel analytical tools that have advanced our mathematical understanding of deep neural networks, and have the potential to help make these “black boxes” more transparent.

This course will introduce:

  • mathematical methods for neural network analysis, providing a solid overview of the analytical tools available to understand neural network models;
  • key connectionist models with links to experimental observations, which provide targets for analytical results.

During the course, you will:

  • attend lectures given by leading researchers on theoretical methods and applications, key connectionist models, and experimental observations;
  • participate in tutorials, Q&A sessions, and panel discussions;
  • present to and engage with lecturers, organisers, and other participants during a poster session;
  • work in a group with other participants on a novel research project, mentored by the course organisers and lecturers.

The course will run full-day Mondays-Fridays and end with a 1.5-day workshop, during which you will hear about current state-of-the-art and the limits of our understanding.

This school offers an in-depth exploration of both theoretical and practical aspects of cognition, neural networks, and machine learning. It will focus on analytical models to understand neural network, with connections to real-world machine learning challenges. Key topics include connectionism and how neural networks simulate human cognition, memory, and learning processes, as well as the intersection of computational neuroscience and decision-making.

Topic lectures will cover the role of memory and learning in the brain, insights into large language models (LLMs) and their application to language tasks, and the cognitive mechanisms behind multitasking and decision-making. The course will also examine the neural mechanisms of decision-making and the cognitive neuroscience of language and semantic memory, offering a comprehensive overview of foundational theories and cutting-edge research in cognition and machine learning.

2

Dates

All deadlines anywhere on earth (AoE)
MilestoneDate
Applications open 4 Apr 2023
Application deadline 15 May 2023
Outcome communicated 7 Jul 2023
School begins28 Aug 2023
3

Schedule

Lecture Presentations Social Idle All timesLondon · BST
Week 1
28 Aug – 1 Sep 2023
Time Mon 8/28 Tue 8/29 Wed 8/30 Thu 8/31 Fri 9/1
09:00 Registration
09:20 Welcome
09:30–10:30 Lecture Sompolinsky Lecture Sompolinsky Lecture Sompolinsky Lecture Sompolinsky Lecture Krzakala
10:30–11:00 break
11:00–12:00 Lecture Sompolinsky Lecture Sompolinsky Synthesis Saxe Lecture Krzakala Lecture Krzakala
12:00–13:30 lunch
13:30–14:30 Lecture McClelland Lecture Krzakala Lecture Krzakala Lecture Krzakala Lecture McClelland
14:45–15:45 Lecture McClelland Lecture McClelland Lecture Krzakala Lecture McClelland Lecture McClelland
15:45–16:00 break
16:00–17:30 Tutorial Satchel Poster session Tutorial Satchel Tutorial Satchel Project work
Week 2
4 – 7 Sep 2023
Time Mon 9/4 Tue 9/5 Wed 9/6 Thu 9/7
09:30–10:30 Panel Lecture Musslick Lecture Musslick Lecture Rogers
10:30–11:00 break
11:00–12:00 Panel Lecture Akrami Lecture Eckstein Lecture Rogers
12:00–13:30 lunch
13:30–14:30 Project work Lecture Akrami Lecture Eckstein
14:45–15:45 Project work Project work Project work
15:45–16:00 break
16:00–17:30 Project presentations
4

Lecturers

5

Participants

List 36 participants
Name
Máté Aller
Jan Philipp Bauer
Ari Benjamin
Anna-Lea Beyer
Victoria Bosch
Abdulkadir Canatar
Elise Chang
Brandon Chen
Chi-Ning Chou
Zach Cohen
Marianne de Heer Kloots
Tala Fakhoury
Dirk Goldschmitt
Mani Hamidi
Jerome Han
Michael Hanna
Eghbal Hosseini
Jaedong Hwang
Akshay Kumar Jagadish
Hajer Karoui
Jin Lee
Xiaoxuan Lei
Huidi Li
Maximilian Mittenbühler
Turan Orujlu
Mitchell Ostrow
Alexandra Proca
Safura Rashid Shomali
Joséphine Raugel
Jirko Rubruck
Kai Sandbrink
Quilee Simeon
Ábel Ságodi
Sushrut Thorat
Elia Turner
Sven Wientjes
Posters 22 posters
Efficiency and (lack of) flexibility in a deep learning model of human spoken word recognition Máté Aller
Quantifying rich and robust inductive biases in chaotic recurrent neural networks Jan Philipp Bauer
The relationship between behavioural tasks and brain space Anna-Lea Beyer
The brain can’t copy-paste: End-to-end topographic neural networks as a way forward for modelling cortical map formation and behaviour Victoria Bosch
Probing biological and artificial neural networks with task-dependent neural manifolds Chi-Ning Chou
What components of NLP models drive similarity to brain activity in language processing? Layer- and head-level analyses Marianne de Heer Kloots
Using representation-learning to guide efficient exploration Mani Hamidi
Understanding subject-verb agreement in pre-trained language models: A circuits approach Michael Hanna
Teasing apart the representational spaces of ANN language models to discover key axes of model-to-brain alignment Eghbal Hosseini
Efficient exploration via fragmentation and recall Jaedong Hwang
Using large-language models to meta-learn human inductive biases Akshay Kumar Jagadish
Human resource-rational planning: A neural network approach Maximilian Mittenbühler
VividDreamer: Tokenized world model with stochastic attention Turan Orujlu
Beyond geometry: Comparing the temporal structure of computation in neural circuits with dynamic mode representational similarity analysis Mitchell Ostrow
Synergistic information supports modality integration and flexible learning in neural networks solving multiple tasks Alexandra Proca
Revealing hidden neuronal microcircuits from correlations among spiking neurons Safura Rashid Shomali
Learning dynamics of semantic knowledge in humans and neural networks Jirko Rubruck
Dimensionality and dynamics of abstract representations Quilee Simeon
An interpretable language for robust neural computation Ábel Ságodi
Characterising representation dynamics in recurrent neural networks for object recognition Sushrut Thorat
The simplicity bias in multi-task RNNs: Shared attractors, reuse of dynamics, and geometric representation Elia Turner
Strategic cognitive control is bound to representations of temporal context Sven Wientjes
6

Organisers

7

Practicalities

Target audience

This course is appropriate for graduate students, postdoctoral fellows and early-career faculty in a number of fields, including psychology, neuroscience, physics, computer science, and mathematics. Attendees are expected to have a strong background in one of these disciplines and to have made some effort to introduce themselves to a complementary discipline.

The course is limited to no more than 40 participants, who will be chosen to balance the representation of different fields. In circumstances where all other things are equal, priority will be given to applicants from underrepresented groups in STEM fields, using positive action under the UK Equality Act 2010 where appropriate.

Course fees

There are no course fees, but attendees are expected to cover their own travel, accommodation and subsistence expenses.

Financial assistance may be available for successful applicants who find it difficult to take up a place for financial reasons. If funding becomes available, successful applicants will be asked to complete a financial aid request form if they need assistance. The amount of financial aid available will depend on the course funding from grants and sponsors.