The rising popularity of intelligent mobile devices and the daunting computational cost of deep learning-based models call for efficient and accurate on-device inference schemes. We propose a quantization scheme that allows inference to be carried out using integer-only arithmetic, which can be implemented more efficiently than floating point inference on commonly available integer-only hardware. We also co-design a training procedure to preserve end-to-end model accuracy post quantization. As a result, the proposed quantization scheme improves the tradeoff between accuracy and on-device latency. The improvements are significant even on MobileNets, a model family known for run-time efficiency, and are demonstrated in ImageNet classification and COCO detection on popular CPUs.

The encoding of sensory information in the human brain is thought to be optimised by two principal processes: ‘prediction’ uses stored information to guide the interpretation of forthcoming sensory events, and ‘attention’ prioritizes these events according to their behavioural relevance. Despite the ubiquitous contributions of attention and prediction to various aspects of perception and cognition, it remains unknown how they interact to modulate information processing in the brain. A recent extension of predictive coding theory suggests that attention optimises the expected precision of predictions by modulating the synaptic gain of prediction error units. Because prediction errors code for the difference between predictions and sensory signals, this model would suggest that attention increases the selectivity for mismatch information in the neural response to a surprising stimulus. Alternative predictive coding models propose that attention increases the activity of prediction (or ‘representation’) neurons and would therefore suggest that attention and prediction synergistically modulate selectivity for ‘feature information’ in the brain. Here, we applied forward encoding models to neural activity recorded via electroencephalography (EEG) as human observers performed a simple visual task to test for the effect of attention on both mismatch and feature information in the neural response to surprising stimuli. Participants attended or ignored a periodic stream of gratings, the orientations of which could be either predictable, surprising, or unpredictable. We found that surprising stimuli evoked neural responses that were encoded according to the difference between predicted and observed stimulus features, and that attention facilitated the encoding of this type of information in the brain. These findings advance our understanding of how attention and prediction modulate information processing in the brain, as well as support the theory that attention optimises precision expectations during hierarchical inference by increasing the gain of prediction errors.

Predictive coding theories argue that recent experience establishes expectations in the brain that generate prediction errors when violated. Prediction errors provide a possible explanation for repetition suppression, where evoked neural activity is attenuated across repeated presentations of the same stimulus. The predictive coding account argues repetition suppression arises because repeated stimuli are expected, whereas non-repeated stimuli are unexpected and thus elicit larger neural responses. Here, we employed electroencephalography in humans to test the predictive coding account of repetition suppression by presenting sequences of visual gratings with orientations that were expected either to repeat or change in separate blocks of trials. We applied multivariate forward modelling to determine how orientation selectivity was affected by repetition and prediction. Unexpected stimuli were associated with significantly enhanced orientation selectivity, whereas selectivity was unaffected for repeated stimuli. Our results suggest that repetition suppression and expectation have separable effects on neural representations of visual feature information.