Visual number sense in untrained deep neural networks

Kim, GwangSu; Jang, Jaeson; Baek, Seungjun; Song, Min; Paik, Se-Bum

doi:10.1126/sciadv.abd6127

Cited by 54 publications

(98 citation statements)

References 59 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Corresponding to the behavior, number neurons respond less when the crows err in the task compared with when they perform accurately [39,42]. Number neurons in the NCL are also spontaneously present in crows even without specific number training [43], a finding suggesting that the neuronal mechanisms for number extraction are inherent to NCL circuits [44,45]. In contrast to NCL neurons, neurons in the crow hippocampus are unresponsive to numbers [46].…”

Section: Evolution Of the Vertebrate Telencephalon (Endbrain)mentioning

confidence: 99%

“…Irrespective of which territory evolves into the highest integration center (the ventral pallium in birds versus the dorsal pallium in mammals), something about the circuit properties of the pallium invites the representation of abstract categories and concepts [44,45]. Among the strategic benefits of the pallium is its role as an intermediary between (thalamic) sensory input and (volitional) motor output.…”

Section: Concluding Remarks and Future Directionsmentioning

confidence: 99%

See 1 more Smart Citation

The Evolutionary History of Brains for Numbers

Nieder

2021

Trends in Cognitive Sciences

View full text Add to dashboard Cite

Humans and other animals share a number sense', an intuitive understanding of countable quantities. Having evolved independent from one another for hundreds of millions of years, the brains of these diverse species, including monkeys, crows, zebrafishes, bees, and squids, differ radically. However, in all vertebrates investigated, the pallium of the telencephalon has been implicated in number processing. This suggests that properties of the telencephalon make it ideally suited to host number representations that evolved by convergent evolution as a result of common selection pressures. In addition, promising candidate regions in the brains of invertebrates, such as insects, spiders, and cephalopods, can be identified, opening the possibility of even deeper commonalities for number sense. A Sense of Number Emerging from Phylogenetically Diverse BrainsHumans and nonhuman animals across the animal kingdom share an intuitive understanding of countable quantities, or numerical quantity (see Glossary) (henceforth 'number') [1]. Distinct zoological groups arising from a common ancestral bilaterian (Box 1)from vertebrates like fishes, birds, and mammals to invertebrates such as insects, spiders, and cephalopodsall possess the ability to discriminate numerical quantity. Numerical quantity refers to the number of elements in a set ('numerosity') and needs to be differentiated from a representation of continuous quantity (Box 2). This 'sense of number' provides survival and reproductive advantages in various animal taxa that can be traced through evolutionary history [2]. HighlightsDiverse species from across the animal tree of life share an intuitive understanding of countable quantities.

show abstract

Section: Evolution Of the Vertebrate Telencephalon (Endbrain)mentioning

confidence: 99%

Section: Concluding Remarks and Future Directionsmentioning

confidence: 99%

The Evolutionary History of Brains for Numbers

Nieder

2021

Trends in Cognitive Sciences

View full text Add to dashboard Cite

show abstract

“…The resulting monotonic responses to numerosity are spatially selective (22), but responses are almost invariant to item size and spacing without the need for explicit object individuation or size normalization (19). Another class of neural network model, deep convolutional neural networks, also show monotonic and numerosity-tuned units, even in randomly-weighted networks (24). Here monotonic units emerge early in the network and feed into numerosity-tuned units, where different weights on these inputs give numerosity-tuned responses with different numerosity preferences.…”

Section: Introductionmentioning

confidence: 99%

Numerosity tuning in human association cortices and local image contrast representations in early visual cortex

Paul

et al. 2021

Preprint

View full text Add to dashboard Cite

Many animals use visual numerosity, the number of items in a group, to guide behavior. Neurons in human association cortices show numerosity-tuned responses, decreasing amplitude with distance from a specific numerosity. How are such responses derived from early visual responses? Recent studies show aggregate response amplitudes in human early visual cortex monotonically increase with numerosity, regardless of object size and spacing. This is surprising because numerosity is typically considered a high-level visual or cognitive feature. Here we first use computational modelling of 7T fMRI data to show these monotonic responses originate at the stimulus's retinotopic location in primary visual cortex (V1). Given this location, we then ask whether these monotonic responses can be better described by V1's established response properties. We characterize the Fourier decomposition (into contrast at specific orientations and spatial frequencies) of laboratory numerosity stimuli. This demonstrates that aggregate Fourier power (at all orientations and spatial frequencies) nonlinearly follows numerosity with little effect of item size, spacing or shape: it is proportional to numerosity at a fixed contrast. This nonlinear relationship lets us distinguish predictions of responses to Fourier power and numerosity. Monotonic responses are better predicted by Fourier power, later tuned responses are better predicted by numerosity. Tuned responses emerge after lateral occipital cortex and are independent of retinotopic location. We propose that numerosity's straightforward perception and neural responses reflect its straightforward estimation from early visual spatial frequency domain image representations. Our numerical vision may have built on behaviorally beneficial analysis of spatial frequency in simpler animals.

show abstract

“…Artificial intelligence (AI) and cognitive science separately investigate how machines and brains solve complex information processing problems, such as language comprehension and visual object recognition. As the artificial neural network approach in AI was in part inspired by biological neural networks, there is continuing interest in comparing the performance of artificial and biological neural networks [1][2][3][4][5][6][7][8][9][10][11] . If humanor animal-like behaviors emerge in artificial neural networks, it indicates that the computations implemented in artificial neural networks can serve as a possible model for human/animal cognition [12][13][14][15] .…”

Section: Introductionmentioning

confidence: 99%

Deep Neural Networks Evolve Human-like Attention Distribution during Goal-directed Reading Comprehension

Zou¹,

Ding²

2021

Preprint

View full text Add to dashboard Cite

Attention is a key mechanism for information selection in both biological brains and many state-of-the-art deep neural networks (DNNs). Here, we investigate whether humans and DNNs allocate attention in comparable ways when seeking information in a text passage to answer a question. We analyze 3 transformer-based DNNs that reach human-level performance when trained to perform the reading comprehension task. We find that the DNN attention distribution quantitatively resembles human attention distribution measured by eye tracking: Human readers fixate longer on words that are more relevant to the question-answering task, demonstrating that attention is modulated by the top-down reading goal, on top of lower-level visual layout and textual features. Further analyses reveal that the attention weights in DNNs are also influenced by both the top-down reading goal and lower-level textual features, with the shallow layers more strongly influenced by lower-level textual features and the deep layers attending more to task-relevant words. Additionally, deep layers’ attention to task-relevant words gradually emerges when pre-trained DNN models are fine-tuned to perform the reading comprehension task, which coincides with the improvement in task performance. These results demonstrate that DNNs can naturally evolve human-like attention distribution through task optimization. The results suggest that human attention during goal-directed reading comprehension is a consequence of task optimization and the attention weights in DNN are of biological significance.

show abstract

Visual number sense in untrained deep neural networks

Cited by 54 publications

References 59 publications

The Evolutionary History of Brains for Numbers

The Evolutionary History of Brains for Numbers

Numerosity tuning in human association cortices and local image contrast representations in early visual cortex

Deep Neural Networks Evolve Human-like Attention Distribution during Goal-directed Reading Comprehension

Contact Info

Product

Resources

About