Extreme neural machines

Recurrent neural networks can solve a variety of computational tasks and produce patterns of activity that capture key properties of brain circuits. However, learning rules designed to train these models are time-consuming and prone to inaccuracies when tuning connection weights located deep within the network. Here, we describe a rapid one-shot learning rule to train recurrent networks composed of biologically-grounded neurons. First, inputs to the model are compressed onto a smaller number of recurrent neurons. Then, a non-iterative rule adjusts the output weights of these neurons based on a target signal. The model learned to reproduce natural images, sequential patterns, as well as a high-resolution movie scene. Together, results provide a novel avenue for one-shot learning in biologically realistic recurrent networks and open a path to solving complex tasks by merging brain-inspired models with rapid optimization rules.

Where and when to look: Sequential effects at the millisecond level

Learning and imitating a complex motor action requires to visually follow complex movements, but conscious perception seems too slow for such tasks. Recent findings suggest that visual perception has a higher temporal resolution at an unconscious than at a conscious level. Here we investigate whether high-temporal resolution in visual perception relies on prediction mechanisms and attention shifts based on recently experienced sequences of visual information. To that aim we explore sequential effects during four different simultaneity/asynchrony discrimination tasks. Two stimuli are displayed on each trial with varying stimulus onset asynchronies (SOA). Subjects decide whether the stimuli are simultaneous or asynchronous and give manual responses. The main finding is an advantage for different-order over same-order trials, when subjects decided that stimuli had been simultaneous on Trial t - 1 , and when Trial t is with an SOA slightly larger than Trial t - 1, or equivalent. The advantage for different-order trials disappears when the stimuli change eccentricity but not direction between trials (Experiment 2), and persists with stimuli displayed in the centre and unlikely to elicit a sense of direction (Experiment 4). It is still observed when asynchronies on Trial t - 1 are small and undetected (Experiment 3). The findings can be explained by an attention shift that is precisely planned in time and space and that incidentally allows subjects to detect an isolated stimulus on the screen, thus helping them to detect an asynchrony.