While people are faster to categorize objects at an intermediate or basic level of specificity (e.g. “bird”), severalrecent studies have shown them to have much earlier access to more general category information (e.g. “animal”). Ultra-rapidsuperordinate classification has been taken as evidence that recognition processes are largely feed-forward. In simulations witha deep neural network model, we show that this conclusion does not follow: even a model that is fully recurrent and interactiveshows ultra-rapid superordinate classification patterns when tested with analogs of behavioral tasks such as rapid serial visualpresentation or deadline classification. Moreover, this recurrent model explains recently-observed similarities and differencesin the time-course of classification as estimated by electro-encephlography (EEG) versus human electro-corticography (ECoG),and also account for the well-known basic-level advantage in non-speeded classification. These results provide evidence thatultra-rapid and unconstrained visual object recognition is supported by interactive processes in the brain.

Generalization to new tasks requires learning of task representations that accurately reflect the similarity structure of the task space. Here, we argue that episodic memory (EM) plays an essential role in this process by stabilizing task representations, thereby supporting the accumulation of structured knowledge. We demonstrate this using a neural network model that infers task representations that minimize the current task's objective function; crucially, the model can retrieve previously encoded task representations from EM and use these to initialize the task inference process. With EM, the model succeeds in learning the underlying task structure; without EM, task representations drift and the network fails to learn the structure. We further show that EM errors can support structure learning by promoting the activation of similar task representations in tasks with similar sensory inputs. Overall, this model provides a novel account of how EM supports the acquisition of structured task representations.

Shared structure in neural responses across people can be obscured because these neural responses sit on different ”co-ordinate systems”; hyperalignment can recover this shared structure by placing different people’s brain responses intoa common functional space (Chen et al., 2015; Haxby et al., 2011). Here, we apply this framework to understand thehidden representations of neural networks. Different neural networks can represent the same input-output mapping usingvery different weights. We show that hyperalignment can construct a shared representational space that recovers sharedrepresentation structure across neural networks. We formally connect representational similarity analysis and hyperalign-ment and use simulations to demonstrate the robustness of hyperalignment against several types of transformations thatpreserve the representation geometry of the network. We also empirically tested our method on some supervised learningbenchmarks (CIFAR10, MNIST) for both standard and convolutional networks.