Publications

Some recent publications on a range of topics.

Relational Convolutional Networks: A framework for learning representations of hierarchical relations

Awni Altabaa and John Lafferty

Compositionality is essential to the success of deep representation learning.
We propose relational convolutional networks as a compositional framework for learning hierarchical relational representations.

Project page

Images with harder-to-reconstruct visual representations leave stronger memory traces

Qi Lin, Zifan Li, John Lafferty, Yilker Ildirim, Nat. Human Behav., 2024

We study a link between memorability and the computation required to approximate an image or visual scene.

Link to paper

The relational bottleneck as an inductive bias for efficient abstraction

Webb et al., Trends in Cog. Sci., 2024

A general inductive bias for abstraction is described and illustrated in several different architectures for deep learning.

Link to paper

Abstractors and relational cross-attention: An inductive bias for explicit relational reasoning in Transformers

Awni Altabaa, Taylor Webb, Jonathan Cohen, and John Lafferty
12th International Conference on Learning Representations (ICLR), Apr 2023

A deep architecture for relational learning is proposed that supports abstraction through a novel attention mechanism.

Project page

Emergent organization of receptive fields in networks of excitatory and inhibitory neurons

Leon Lufkin, Ashish Puri, Ganlin Song, Xinyi Zhou, John Lafferty

Sparse coding algorithms with local patterns of excitatory and inhibitory connections lead to receptive fields that exhibit patterns characteristic of V1, and show patterns of word semantics when applied to word embeddings in language modeling.

Link to paper

Shallow neural networks trained to detect collisions recover features of visual loom-sensitive neurons

Baohua Zhou, Zifan Li, Sunnie Kium, John Lafferty, Damon Clark

eLife, 2022

Artificial neural networks with biologically informed architecture learn receptive fields that resembles thos of the compound insect eye, when trained on synthetic videos of looming objects.

Link to paper

Convergence and alignment of gradient descent with random backpropagation weights

Ganlin Song, Ruitu Xu, John Lafferty

NeurIPS 2021

We prove convergence of the feedback alignment algorithm for two-layer networks, proposed as a biologically plausible alternative to backpropgation, and show that alignment requires regularization.

Link to paper