Fig 1.
‘Explaining away’ in sensory perception.
(a) The presumed goal of perception is to infer the state of the external world from received sensory cues. Here, two possible events (someone arriving at the door, and a telephone call) can give rise to three sensory cues (a knocking sound, ringing sound, or vibration). The ringing sound is ambiguous: it can come from either the door bell or the phone. Cues, such as a vibrating telephone, can resolve this ambiguity: here, increasing the chances that the phone is ringing, while decreasing the chances that there is someone at the door. Such competition between different explanations for received sensory cues is called ‘explaining away’. (b-c) In sensory neural circuits, explaining away results in suppression from non-preferred stimuli in the surround. Its effects vary dramatically, depending on whether inhibition acts (b) globally on the neural responses or (c) selectively, on certain neural inputs. (d-e) Hypothetical response of ‘door’ and ‘phone’ selective neurons, in response to different combinations of sensory cues. The qualitative effects of explaining away depend on whether it (d) globally suppresses the response of one or other detector, or (e) selectively suppresses the influence of certain cues.
Fig 2.
Input-targeted inhibition alters neural selectivity.
(a) Schematic of neural network, with input-targeted feedback. (b) Steady-state response responses of recorded neuron, predicted by model with input-targeted divisive feedback, or subtractive inhibition. There are three stimulus conditions: (i) ‘no context’ condition, with a single stimulus within the cell’s RF; (ii) ‘adjoint context’ condition, with a second stimulus in the surround, near to the stimulus within the RF and (iii) ‘disjoint context’ condition, with a second stimulus in the surround, far from the stimulus within the RF. (c) Contextual shifts in neural tuning curves. Each neuron encodes a stimulus features (e.g. orientation, or motion direction) with a given preferred value. The mean response of a single neuron is plotted against the presented stimulus value, in the absence (black) or presence of an overlapping mask, to the right or left of the neuron’s preferred stimulus (blue and red). (c, lower panel) As above, but for an LN model. (d) Simulation of network in which cells encode stimuli in a circular region of space. (top panel) Estimated RF, with random sparse stimulus. (lower panel) Estimated RF in presence of vertical mask. The measured RF is elongated in the horizontal direction. (e) As for panel d, but for an LN model.
Fig 3.
(a) Contextual reshaping of multimodal RFs. Each neuron encodes a stimulus feature (e.g. an odor) that is assumed to elicit a multimodal pattern of sensory activity (upper panels). Neural RFs are measured in the presence of a mask stimulus that activates a small number of nearby receptors. For the three cells shown, recorded RFs undergo complex, non-local changes in the presence of the contextual mask. (b) Reshaping of neural RFs in a simplified network of three neurons, which encode the letters ‘V’, ‘I’, and ‘A’. (c) The RF of a neuron encoding the letter ‘I’ is significantly altered by a contextual stimulus designed to selectively activate one of the other two neurons in the network.
Fig 4.
(a) Schematic illustrating how contextual shifts in neural tuning curves required for a context-invariant neural code. (b) Contextual mask presented in each condition. (c, left panel) Tuning curve of a model neuron in the presence of the three different stimulus masks (tuning curves are rescaled, to have zero mean and unitary standard-deviation). (right panel) Inferred readout filters for the same neuron in each condition. (d) As for panel c, but for an LN model. (e) Mean squared difference in (rescaled) tuning curves across the different stimulus contexts. Each cell corresponds to one data point. The example cell, plotted in panels c-d is shown in red. (f) Identical analysis to panel e, but applied to the linear readout filters. (g) Normalized reconstruction error using ‘correct’ readout filters for each stimulus condition (blue bars), or ‘mismatched’ decoders, inferred in other stimulus conditions.
Fig 5.
Input-targeted inhibition allows for discrimination of similar stimulus features.
(a) Three different stimulus features (e.g. odors) encoded by different neurons in the network. The plots show the overlapping pattern of receptor activation elicited by each feature. (b) Three different combinations of features presented to the network. (c) Neural responses to each feature combination, obtained from the input-targeted divisive inhibition model. The response of each neuron is highly specific to its encoded feature, even with multiple overlapping features presented simultaneously. (d) As for panel c, but with an LN model, trained to match the responses of the divisive input model, to a range of different presented feature combinations. In contrast to before, neurons respond non-specifically when similar featue are presented together.
Fig 6.
Proposed neural implementation.
(a) Example network in which each stimulus feature is encoded by an excitatory neuron that projects to higher level areas. Divisive inhibition acts on individual synaptic inputs. (b) Example network with two neural populations: excitatory neurons encode the ratio between the received and predicted input, , while inhibitory neurons encode estimated stimulus features,
. (c) Example of a hierarchical network. The fractional prediction error encoded in a given layer is integrated by downstream neurons, which encode more complex stimulus features. (d) Divisive gain control. (left) A ‘test’ stimulus activates the input to the recorded neuron (indicated with arrow), while a mask stimulus activates the input to the other neuron. Response of recorded neuron is plotted versus amplitude of the test stimulus. Each plot corresponds to a different amplitude mask (see legend).
Fig 7.
Predicted temporal response profile.
(a, left) Temporal response profile of excitatory neuron, to a constant feed-forward input of varying strength. (right) Instantaneous response of the excitatory neuron versus amplitude of feed-forward input. Each plot corresponds to a fixed time after stimulus presentation (indicated by vertical dashed lines in left panel). (b) Same as (a), but for an inhibitory model neuron. (c-d) Same as a-b, but for a model with subtractive, rather than divisive inhibition.
Fig 8.
Traveling waves in the visual cortex.
(a) Schematic of topographic model network, in which each inhibitory neuron connects with equal strength to two neighbouring excitatory neurons. The feed-forward input decreases with distance from the centre. (b) Heat map of excitatory neural responses, normalized by peak amplitude. Each row shows the response of a neuron at a specified distance from the centre. Filled and solid diamonds indicate at what time each neuron’s response is 70% of its maximum. The right panel indicates the maximum response of each neuron. (c) Same as c, but with subtractive, rather than divisive inhibition.