Sushrut Thorat

Sushrut Thorat, Adrien Doerig, Tim C. Kietzmann

Recurrent neural networks (RNNs) have yielded promising results for both recognizing objects in challenging conditions and modeling aspects of primate vision. However, the representational dynamics of recurrent computations remain poorly understood, especially in large-scale visual models. Here, we studied such dynamics in RNNs trained for object classification on MiniEcoset, a novel subset of ecoset. We report two main insights. First, upon inference, representations continued to evolve after correct classification, suggesting a lack of the notion of being ``done with classification''. Second, focusing on ``readout zones'' as a way to characterize the activation trajectories, we observe that misclassified representations exhibit activation patterns with lower L2 norm, and are positioned more peripherally in the readout zones. Such arrangements help the misclassified representations move into the correct zones as time progresses. Our findings generalize to networks with lateral and top-down connections, and include both additive and multiplicative interactions with the bottom-up sweep. The results therefore contribute to a general understanding of RNN dynamics in naturalistic tasks. We hope that the analysis framework will aid future investigations of other types of RNNs, including understanding of representational dynamics in primate vision.

Daniel Anthes, Sushrut Thorat, Peter König et al.

Unlike primates, training artificial neural networks on changing data distributions leads to a rapid decrease in performance on old tasks. This phenomenon is commonly referred to as catastrophic forgetting. In this paper, we investigate the representational changes that underlie this performance decrease and identify three distinct processes that together account for the phenomenon. The largest component is a misalignment between hidden representations and readout layers. Misalignment occurs due to learning on additional tasks and causes internal representations to shift. Representational geometry is partially conserved under this misalignment and only a small part of the information is irrecoverably lost. All types of representational changes scale with the dimensionality of hidden representations. These insights have implications for deep learning applications that need to be continuously updated, but may also aid aligning ANN models to the rather robust biological vision.

Ishita Darade, Sushrut Thorat

Structured prompts require integrating components according to task-relevant relations. How a network implements this integration is often hard to judge in language or vision, where those relations are rarely specified precisely enough to define a candidate internal algorithm. Arithmetic offers a cleaner setting. We study a Transformer trained on base-digit extraction: given $N$, $B$, and $D$, it must report the coefficient of $B^D$ in the base-$B$ expansion of $N$. The closed-form solution, $\lfloor N/B^D \rfloor \bmod B$, provides explicit candidate algorithmic intermediates. Across three seeds, the model reaches 99.83% exact-answer accuracy on held-out number-base intersections, establishing reliable task competence. Linear probes decode the intermediates, making staged arithmetic computation plausible. Causal tests then separate representation from use: within the localized route from the stream with $D$ as input to the output positions, behavior depends on early $D$-selective communication, independent of $N$ and $B$. Relatedly, a sparse circuit search finds mostly separate $N$, $B$, and $D$ routes that combine late rather than the staged route suggested by the probes. Thus, the model represents the intermediates that make the closed-form solution plausible, but the identified localized causal route does not transmit them to the output stream. This case shows that probe-based conclusions can diverge sharply from causal observations, even when explicit algorithmic hypotheses are available.

Daniel Anthes, Sushrut Thorat, Peter König et al.

Continual learning algorithms strive to acquire new knowledge while preserving prior information. Often, these algorithms emphasise stability and restrict network updates upon learning new tasks. In many cases, such restrictions come at a cost to the model's plasticity, i.e. the model's ability to adapt to the requirements of a new task. But is all change detrimental? Here, we approach this question by proposing that activation spaces in neural networks can be decomposed into two subspaces: a readout range in which change affects prior tasks and a null space in which change does not alter prior performance. Based on experiments with this novel technique, we show that, indeed, not all activation change is associated with forgetting. Instead, only change in the subspace visible to the readout of a task can lead to decreased stability, while restricting change outside of this subspace is associated only with a loss of plasticity. Analysing various commonly used algorithms, we show that regularisation-based techniques do not fully disentangle the two spaces and, as a result, restrict plasticity more than need be. We expand our results by investigating a linear model in which we can manipulate learning in the two subspaces directly and thus causally link activation changes to stability and plasticity. For hierarchical, nonlinear cases, we present an approximation that enables us to estimate functionally relevant subspaces at every layer of a deep nonlinear network, corroborating our previous insights. Together, this work provides novel means to derive insights into the mechanisms behind stability and plasticity in continual learning and may serve as a diagnostic tool to guide developments of future continual learning algorithms that stabilise inference while allowing maximal space for learning.

Linda Ariel Ventura, Victoria Bosch, Tim C Kietzmann et al.

Adaptive cognition requires structured internal models representing objects and their relations. Predictive neural networks are often proposed to form such "world models", yet their underlying mechanisms remain unclear. One hypothesis is that action-conditioned sequential prediction suffices for learning such world models. In this work, we investigate this possibility in a minimal in-silico setting. Sequentially sampling tokens from 2D continuous token scenes, a recurrent neural network is trained to predict the upcoming token from current input and a saccade-like displacement. On novel scenes, prediction accuracy improves across the sequence, indicating in-context learning. Decoding analyses reveal path integration and dynamic binding of token identity to position. Interventional analyses show that new bindings can be learned late in sequence and that out-of-distribution bindings can be learned. Together, these results demonstrate how structured representations that rely on flexible binding emerge to support prediction, offering a mechanistic account of sequential world modeling relevant to cognitive science.

Zejin Lu, Sushrut Thorat, Radoslaw M Cichy et al.

Despite years of research and the dramatic scaling of artificial intelligence (AI) systems, a striking misalignment between artificial and human vision persists. Contrary to humans, AI heavily relies on texture-features rather than shape information, lacks robustness to image distortions, remains highly vulnerable to adversarial attacks, and struggles to recognise simple abstract shapes within complex backgrounds. To close this gap, we here introduce a solution that arises from a previously underexplored direction: rather than scaling up, we take inspiration from how human vision develops from early infancy into adulthood. We quantified the visual maturation by synthesising decades of psychophysical and neurophysiological research into a novel developmental visual diet (DVD) for AI vision. We show that guiding AI systems through this human-inspired curriculum produces models that closely align with human behaviour on every hallmark of robust vision tested yielding the strongest reported reliance on shape information to date, abstract shape recognition beyond the state of the art, higher robustness to image corruptions, and stronger resilience to adversarial attacks. By outperforming high parameter AI foundation models trained on orders of magnitude more data, we provide evidence that robust AI vision can be achieved by guiding the way how a model learns, not merely how much it learns, offering a resource-efficient route toward safer and more human-like artificial visual systems.

Victoria Bosch, Daniel Anthes, Adrien Doerig et al.

Large language models (LLMs) have revolutionized human-machine interaction, and have been extended by embedding diverse modalities such as images into a shared language space. Yet, neural decoding has remained constrained by static, non-interactive methods. We introduce CorText, a framework that integrates neural activity directly into the latent space of an LLM, enabling open-ended, natural language interaction with brain data. Trained on fMRI data recorded during viewing of natural scenes, CorText generates accurate image captions and can answer more detailed questions better than controls, while having access to neural data only. We showcase that CorText achieves zero-shot generalization beyond semantic categories seen during training. Furthermore, we present a counterfactual analysis that emulates in-silico cortical microstimulation. These advances mark a shift from passive decoding toward generative, flexible interfaces between brain activity and language.

Rowan P. Sommers, Sushrut Thorat, Daniel Anthes et al.

Flexible cognition demands discovering hidden rules to quickly adapt stimulus-response mappings. Standard neural networks struggle in such tasks requiring rapid, context-driven remapping. Recently, Hummos (2023) introduced a fast-and-slow learning algorithm to mitigate this shortcoming, but its scalability to complex, image-computable tasks was unclear. Here, we propose the Wisconsin Neural Network (WiNN), which extends Hummos' fast-and-slow learning to image-computable tasks demanding flexible rule-based behavior. WiNN employs a pretrained convolutional neural network for vision, coupled with an adjustable "context state" that guides attention to relevant features. If WiNN produces an incorrect response, it first iteratively updates its context state to refocus attention on task-relevant cues, then performs minimal parameter updates to attention and readout layers. This strategy preserves generalizable representations in the sensory and attention networks, reducing catastrophic forgetting. We evaluate WiNN on an image-based extension of the Wisconsin Card Sorting Task, revealing several markers of cognitive flexibility: (i) WiNN autonomously infers underlying rules, (ii) requires fewer examples to do so than control models reliant on large-scale parameter updates, (iii) can perform context-based rule inference solely via context-state adjustments-further enhanced by slow updates of attention and readout parameters, and (iv) generalizes to unseen compositional rules through context-state updates alone. By blending fast context inference with targeted attentional guidance, WiNN achieves "sparks" of flexibility. This approach offers a path toward context-sensitive models that retain knowledge while rapidly adapting to complex, rule-based tasks.

Sushrut Thorat, Adrien Doerig, Alexander Kroner et al.

Scenes are complex, yet structured collections of parts, including objects and surfaces, that exhibit spatial and semantic relations to one another. An effective visual system therefore needs unified scene representations that relate scene parts to their location and their co-occurrence. We hypothesize that this structure can be learned self-supervised from natural experience by exploiting the temporal regularities of active vision: each fixation reveals a locally-detailed glimpse that is statistically related to the previous one via co-occurrence and saccade-conditioned spatial regularities. We instantiate this idea with Glimpse Prediction Networks (GPNs) -- recurrent models trained to predict the feature embedding of the next glimpse along human-like scanpaths over natural scenes. GPNs successfully learn co-occurrence structure and, when given relative saccade location vectors, show sensitivity to spatial arrangement. Furthermore, recurrent variants of GPNs were able to integrate information across glimpses into a unified scene representation. Notably, these scene representations align strongly with human fMRI responses during natural-scene viewing across mid/high-level visual cortex. Critically, GPNs outperform architecture- and dataset-matched controls trained with explicit semantic objectives, and match or exceed strong modern vision baselines, leaving little unique variance for those alternatives. These results establish next-glimpse prediction during active vision as a biologically plausible, self-supervised route to brain-aligned scene representations learned from natural visual experience.

Sushrut Thorat, Giacomo Aldegheri, Tim C. Kietzmann

Recurrent neural networks (RNNs) have been shown to perform better than feedforward architectures in visual object categorization tasks, especially in challenging conditions such as cluttered images. However, little is known about the exact computational role of recurrent information flow in these conditions. Here we test RNNs trained for object categorization on the hypothesis that recurrence iteratively aids object categorization via the communication of category-orthogonal auxiliary variables (the location, orientation, and scale of the object). Using diagnostic linear readouts, we find that: (a) information about auxiliary variables increases across time in all network layers, (b) this information is indeed present in the recurrent information flow, and (c) its manipulation significantly affects task performance. These observations confirm the hypothesis that category-orthogonal auxiliary variable information is conveyed through recurrent connectivity and is used to optimize category inference in cluttered environments.

Sushrut Thorat, Giacomo Aldegheri, Marcel A. J. van Gerven et al.

In daily life situations, we have to perform multiple tasks given a visual stimulus, which requires task-relevant information to be transmitted through our visual system. When it is not possible to transmit all the possibly relevant information to higher layers, due to a bottleneck, task-based modulation of early visual processing might be necessary. In this work, we report how the effectiveness of modulating the early processing stage of an artificial neural network depends on the information bottleneck faced by the network. The bottleneck is quantified by the number of tasks the network has to perform and the neural capacity of the later stage of the network. The effectiveness is gauged by the performance on multiple object detection tasks, where the network is trained with a recent multi-task optimization scheme. By associating neural modulations with task-based switching of the state of the network and characterizing when such switching is helpful in early processing, our results provide a functional perspective towards understanding why task-based modulation of early neural processes might be observed in the primate visual cortex

Sushrut Thorat, Marcel van Gerven, Marius Peelen

Visual object recognition is not a trivial task, especially when the objects are degraded or surrounded by clutter or presented briefly. External cues (such as verbal cues or visual context) can boost recognition performance in such conditions. In this work, we build an artificial neural network to model the interaction between the object processing stream (OPS) and the cue. We study the effects of varying neural and representational capacities of the OPS on the performance boost provided by cue-driven feature-based feedback in the OPS. We observe that the feedback provides performance boosts only if the category-specific features about the objects cannot be fully represented in the OPS. This representational limit is more dependent on task demands than neural capacity. We also observe that the feedback scheme trained to maximise recognition performance boost is not the same as tuning-based feedback, and actually performs better than tuning-based feedback.

Sushrut Thorat, Varad Choudhari

In this paper, we outline an approach to build graph-based reverse dictionaries using word definitions. A reverse dictionary takes a phrase as an input and outputs a list of words semantically similar to that phrase. It is a solution to the Tip-of-the-Tongue problem. We use a distance-based similarity measure, computed on a graph, to assess the similarity between a word and the input phrase. We compare the performance of our approach with the Onelook Reverse Dictionary and a distributional semantics method based on word2vec, and show that our approach is much better than the distributional semantics method, and as good as Onelook, on a 3k lexicon. This simple approach sets a new performance baseline for reverse dictionaries.