Orthogonal representations for robust context-dependent task performance in brains and neural networks

Flesch, Timo; Juechems, Keno; Dumbalska, Tsvetomira; Saxe, Andrew M.; Summerfield, Christopher

doi:10.1016/j.neuron.2022.01.005

Cited by 134 publications

(173 citation statements)

References 54 publications

Supporting

Mentioning

127

Contrasting

Order By: Relevance

“…5a-c,e). Previous studies have shown that the brain uses private, orthogonal subspaces to disentangle information about different aspects of a task (Mante et al 2013; Kaufman et al 2014; Sheahan, Franklin, and Wolpert 2016; Bagur et al 2018; Flesch et al 2022). For example, population recordings in the premotor cortex during a context-dependent perceptual task show that monkeys use the context cue to “displace” the state of the network such that the encoding of the irrelevant stimulus feature is decoupled from the choice axis (Mante et al 2013).…”

Section: Discussionmentioning

confidence: 99%

Ecologically pre-trained RNNs explain suboptimal animal decisions

Molano-Mazón

Duque

et al. 2021

Preprint

View full text Add to dashboard Cite

When faced with a new task, animals′ cognitive capabilities are determined both by individual experience and by structural priors evolved to leverage the statistics of natural environments. Rats can quickly learn to capitalize on the trial sequence correlations of two-alternative forced choice (2AFC) tasks after correct trials, but consistently deviate from optimal behavior after error trials, when they waive the accumulated evidence. To understand this outcome-dependent gating, we first show that Recurrent Neural Networks (RNNs) trained in the same 2AFC task outperform animals as they can readily learn to use previous trials′ information both after correct and error trials. We hypothesize that, while RNNs can optimize their behavior in the 2AFC task without a priori restrictions, rats′ strategy is constrained by a structural prior adapted to a natural environment in which rewarded and non-rewarded actions provide largely asymmetric information. When pre-training RNNs in a more ecological task with more than two possible choices, networks develop a strategy by which they gate off the across-trial evidence after errors, mimicking rats′ behavior. Our results suggest that the observed suboptimal behavior reflects the influence of a structural prior that, adaptive in a natural multi-choice environment, constrains performance in a 2AFC laboratory task.

show abstract

Section: Discussionmentioning

confidence: 99%

Ecologically pre-trained RNNs explain suboptimal animal decisions

Molano-Mazón

Duque

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

“…A recent study in which human participants learned to perform two tasks while in a functional magnetic resonance (fMRI) scanner provides some evidence for our predictions[9]. The representations of a high-dimensional stimulus with two task-relevant dimensions (one which was relevant in each of two contexts) were studied in both the fMRI imaging data and in neural networks that were trained to perform the two tasks (the setup in this work is similar to certain manipulations in our study, particularly to the partial information case shown in fig.…”

Section: Discussionmentioning

confidence: 76%

“…Further, experimental work in the hippocampus and prefrontal cortex has shown that representations of the sensory and cognitive features related to a complex cognitive task, also support generalization[5]. We refer to representations of task-relevant sensory and cognitive variables that support generalization – like in these examples and others[9, 10] – as abstract representations.…”

Section: Introductionmentioning

confidence: 99%

Abstract representations emerge naturally in neural networks trained to perform multiple tasks

Johnston

Fusi

2021

Preprint

View full text Add to dashboard Cite

Humans and other animals demonstrate a remarkable ability to generalize knowledge across distinct contexts and objects during natural behavior. We posit that this ability depends on the geometry of the neural population representations of these objects and contexts. Specifically, abstract, or disentangled, neural representations -- in which neural population activity is a linear function of the variables important for making a decision -- are known to allow for this kind of generalization. Further, recent neurophysiological studies have shown that the brain has sufficiently abstract representations of some sensory and cognitive variables to enable generalization across distinct contexts. However, it is unknown how these abstract representations emerge. Here, using feedforward neural networks, we demonstrate a simple mechanism by which these abstract representations can be produced: The learning of multiple distinct classification tasks. We demonstrate that, despite heterogeneity in the task structure, abstract representations that enable reliable generalization can be produced from a variety of different inputs -- including standard nonlinearly mixed inputs, inputs that mimic putative representations from early sensory areas, and even simple image inputs from a standard machine learning data set. Thus, we conclude that abstract representations of sensory and cognitive variables emerge from the multiple behaviors that animals exhibit in the natural world, and may be pervasive in high-level brain regions. We make several specific predictions about which variables will be represented abstractly as well as show how these representations can be detected.

show abstract

“…As such, these learning dynamics naturally give rise to low-dimensional neural representations. Such learning dynamics may thus underlie the popular idea in neuroscience of low-dimensional neural manifolds (see Flesch et al (2022)).…”

Section: Discussionmentioning

confidence: 99%

Efficient neural codes naturally emerge through gradient descent learning

Benjamin

Zhang

Qiu

et al. 2022

Preprint

View full text Add to dashboard Cite

Animal sensory systems are more sensitive to common features in the environment than uncommon features. For example, small deviations from the more frequently encountered horizontal orientations can be more easily detected than small deviations from the less frequent diagonal ones. Here we find that artificial neural networks trained to recognize objects also have patterns of sensitivity that match the statistics of features in images. To interpret these findings, we show mathematically that learning with gradient descent in deep neural networks preferentially creates representations that are more sensitive to common features, a hallmark of efficient coding. This result suggests that efficient coding naturally emerges from gradient-like learning on natural stimuli.

show abstract

Orthogonal representations for robust context-dependent task performance in brains and neural networks

Cited by 134 publications

References 54 publications

Ecologically pre-trained RNNs explain suboptimal animal decisions

Ecologically pre-trained RNNs explain suboptimal animal decisions

Abstract representations emerge naturally in neural networks trained to perform multiple tasks

Efficient neural codes naturally emerge through gradient descent learning

Contact Info

Product

Resources

About