Task representations in neural networks trained to perform many cognitive tasks.

Area of Science:

• Computational neuroscience
• Cognitive science
• Artificial intelligence

Background:

• Traditional studies face limitations in elucidating brain mechanisms for multiple tasks.
• Understanding cognitive flexibility requires models capable of handling diverse tasks.

Purpose of the Study:

• To investigate the neural mechanisms underlying cognitive flexibility using computational models.
• To explore how single network models perform multiple cognitive tasks.

Main Methods:

• Trained single recurrent neural network models on 20 diverse cognitive tasks.
• Developed a measure to quantify relationships between neural representations of tasks.
• Employed a continual-learning technique to train networks sequentially.

Main Results:

• Recurrent units formed functionally specialized clusters after training.
• Task representations exhibited compositionality, enabling task recombination.
• Networks displayed mixed task selectivity, mirroring prefrontal neuron activity.

Conclusions:

• Computational models can reveal mechanisms of cognitive flexibility.
• Specialized units and compositional representations are key to performing multiple tasks.
• This work offers a platform for studying neural representations across various cognitive tasks.