Flexible rule learning in machines

With: Rowan Sommers, Daniel Anthes, Tim Kietzmann
Summary: Inspired by Hummos, we built a system to solve an image-based variant of the Wisconsin Card Sorting Task, and found behaviors suggesting sparks of cognitive flexibility. Main finding: compositional rule inference in activity space.
Comments: Preprint

Structured representational drift aids continual learning

With: Daniel Anthes, Peter König, Tim Kietzmann
Summary: In contemporary continual learning, readout misalignment due to learning-induced representational drift poses a big problem. However, constraining this drift to the readout null-space helps networks be both stable and plastic.
Publication: CCN'23 paper, CoLLAs'24 paper
Comments: CCN paper in brief, new analyses @ CCN'24

Attentional Routing is as effective as Direct Access

With: Johannes Singer, Radoslaw Cichy, Tim Kietzmann
Summary: ANN models reveal that attentional routing can push task-relevant information residing in lower-level areas through the network as effectively as directly accessing that information from those areas. These results challenge claims in Birman & Gardner 2019 about the need of direct access in explaining human behavior.
Publication: CCN'24 paper

Assessing the emergence of an attention schema in object tracking

With: Lotta Piefke, Adrien Doerig, Tim Kietzmann
Summary: In tracking an object through clutter with spatial attention, with RL, an agent learns to create an explicit encoding of the attentional state - an attention schema. This schema is most useful when the attentional state cannot be inferred from the stimulus.
Publication: CogSci'24 paper
Comments: CogSci paper in brief

Size-dependence of object search templates in natural scenes

With: Surya Gayet, Marius Peelen, et al.
Summary: Object size varies with the location of the object in scenes. During search for an object, in addition to the object's identity, the attentional template contains information its size, entangled with its identity, which is inferred from its location in the scene.
Publication: JEP:HPP'24 paper
Comments: JEP:HPP paper in brief

Bodies as features in visual search

With: Marius Peelen
Summary: Are high-level visual features prioritised, via feature-based attention, spatially-globally? We found attentional gain modulation of the fMRI representations of body silhouettes, presented in task-irrelevant locations, in high-level visual cortex.
Publication: NeuroImage'22 paper
Comments: NeuroImage paper in brief, Code + Data

The function of early task-based modulations in object detection

With: Giacomo Aldegheri, Marcel van Gerven, Marius Peelen
Summary: Task-based modulation of early visual processing in neural networks alleviates subsequent capacity limits caused by task and neural constraints. Bias/gain modulation of neural activations can be linked to tapping into a superposition of networks. Optimised neural modulations are not feature-similarity gain modulations.
Publications: CCN'18 paper, CCN'19 paper

The influence of scene information on object processing

With: Ilze Thoonen, Sjoerd Meijer, Marius Peelen
Summary: Task-irrelevant scene information biases categorization response towards co-varying objects (e.g. cars on roads). However, no evidence is found, across 4 experiments, for task-irrelevant scene information boosting the sensitivity of detecting co-varying objects. Further experimentation is required for validating these observations.
Comments: Summary presentation

Decision-making RNNs

With: Jonas Bieber, Tim Kietzmann
Summary: Human categorisation reaction times can be inferred from RNN representations/outputs (Spoerer et al. 2020, Goetschalckx et al. 2023). Instead, we explore whether RNNs can generate RTs that are comparable to human RTs.

How does recurrence interact with feedforward processing in RNNs?

With: Adrien Doerig, Tim Kietzmann
Summary: In RNNs performing image classification, the feedforward sweep instantiates a representational arrangement that dovetails with the recurrence-induced "equal movement for all representations" prior, allowing classifications to be corrected.
Publication: CCN'23 paper
Comments: CCN paper in brief

Recurrent operations in neural networks trained to recognise objects

With: Giacomo Aldegheri, Tim Kietzmann
Summary: In a recurrent neural network trained for object categorization, the recurrent flow carries category-orthogonal object feature (e.g. object location) information, which is used, iteratively, to constrain the subsequent inferences about the object's category.
Publication: SVRHM'21 paper
Comments: SVRHM paper in brief

Shape/texture coding in macaque visual cortex

With: Zejin Lu, Daniel Anthes, Tim Kietzmann
Summary: Many studies suggest that the ventral visual stream's responses are mostly driven by texture information. Mostly this is done by distorting the shape elements and assessing the response. Instead, we pit shape against texture (cue-conflict images) and assess which factor drives the response.

Relational representations via glimpse prediction

With: Linda Ventura, Tim Kietzmann, et al.
Summary: Inspired by Summerfield et al. 2020, research on RF remapping, and predictive vision, we evaluate the usefulness of predicting the content of the next glimpse towards generating scene representations that bear relational information about its constituents.

Glimpse prediction for human-like scene representation

With: Adrien Doerig, Tim Kietzmann, et al.
Summary: During eye movements, learning to predict the contents of the next glimpse, given the saccadic efference copy, coaxes a neural network to encode the co-occurrence and spatial arrangement of parts of natural scenes in a code that aligns with representations of natural scene images in the human ventral visual cortex.
Comments: Preliminary results were presented at CCN'25

Representational drift in macaque visual cortex

With: Daniel Anthes, Peter König, Tim Kietzmann
Summary: Employing tools developed during our investigations into continual learning, we study if representational drift occurs in macaque visual cortex and how that multi-area system deals with changing representations.

Brain reading with a Transformer

With: Victoria Bosch, Tim Kietzmann, et al.
Summary: Using fMRI responses to natural scenes to condition the sentence generation in a Transformer, we study the neural underpinnings of scene semantics (objects and their relationships) encoded in natural language.
Comments: Preliminary results were presented at CCN'24

Developmentally-inspired shape bias in artificial neural networks

With: Zejin Lu, Radoslaw Cichy, Tim Kietzmann
Summary: Inspired by the Adaptive Initial Degradation hypothesis, we trained ANNs with a graded coarse-to-fine image diet and found that their classification behavior becomes highly shape-biased! This setup also confers distortion and adversarial robustness.
Comments: Preprint, under review.

Perception of rare inverted letters among upright ones

With: Jochem Koopmans, Genevieve Quek, Marius Peelen
Summary: In a Sperling-like task where the letters are mostly upright, there is a general tendency to report occasionally-present and absent inverted letters as upright to the same extent. Previously reported expectation-driven illusions might be post-perceptual.
Comments: Jochem's masters thesis. Paper under review.

Task-dependent characteristics of neural multi-object processing

With: Lu-Chun Yeh, Marius Peelen
Summary: The association between the neural processing of multi-object displays and the representations of those objects presented in isolation is task-dependent: same/different judgement relates to earlier, and object search to later stages in MEG/fMRI signals.
Publication: JNeurosci'24 paper
Comments: JNeurosci paper in brief

Statistical learning of distractor co-occurrences facilitates visual search

With: Genevieve Quek, Marius Peelen
Summary: Efficient visual search relies on the co-occurrence statistics of distractor shapes. Increased search efficiency among co-occurring distractors is probably driven by faster and/or more accurate rejection of a distractor's partner as a possible target.
Publication: JOV'22 paper
Comments: JOV paper in brief

The nature of the animacy organization in human ventral temporal cortex

With: Daria Proklova, Daniel Kaiser, Marius Peelen
Summary: The animacy organisation in the ventral temporal cortex is not fully driven by visual feature differences (modelled with a CNN). It also depends on non-visual (inferred) factors such as agency (quantified through a behavioural task).
Publications: eLife'19 paper, Masters Thesis
Comments: Masters thesis in brief, eLife paper in brief

Reverse dictionary using a word-definition based graph search

With: Varad Choudhari
Summary: A method to process any forward word dictionary to build a reverse dictionary, using a n-hop reverse search on a graph, through word definitions. Performs as well as the state-of-the-art on a 3k lexicon. Doesn't scale well to 80k.
Publication: COLING'16 Paper
Comments: COLING paper in brief

A Spiking Neural Network as a Quadcopter Flight Controller

With: Sukanya Patil, Bipin Rajendran
Summary: a. Model-based control system for quadcopters towards velocity-waypoint navigation.
b. Modular SNNs for real-time arithmetic operations, using plastic synapses. SNNs are hard to tame!
Publications: IJCNN'15 paper, B.Tech. Thesis
Comments: Thesis rumination, IJCNN paper in brief