| Summary: | <p>According to the theory of predictive coding, the brain does not generate sensations simply by ’sponging up’ bottom-up information arriving from the senses. Instead, cortical circuits are fundamentally wired to actively make predictions on the basis of experience, communicate these in a top-down fashion to areas with access to less global information, and prop- agate the residual errors that ensue when predictions are compared with reality (Friston, 2005). As these error signals feed forward through the processing hierarchy, they allow each successive processing stage to refine its predictions, allowing future stimulation to be better predicted and, in some ways, better ’understood’ by the brain’s internal model of the world. The critical predictive and error signals are thought to stem from anatomically distinct neural populations, with predictions represented both in the deep and superficial cortical layers, and errors localised to superficial sites (Kanai et al., 2015). Evidence that the superficial and deep layers are, to some degree, specialised for bottom-up and top-down communications, affiliated with activity in the γ (30-100 Hz) and α-β (8-25 Hz) frequency bands respectively (Bastos et al., 2015; Michalareas et al., 2016) has also raised the possibility that predictions and errors are encoded in different spectral channels. The goal of this thesis work was to examine the hypothesised anatomical and functional affiliation of predictive coding variables with distinct cortical layers and their oscillatory signatures. This question was pursued on a range of scales - from the behaving animal and isolated in vitro brain circuit, to the single cell.</p>
<p>In the first chapter, I piloted two behavioural paradigms in which mice made 2-Alternative Forced Choice judgements about visual stimuli while auditory cues predicting stimulus orientation were used to incentivise sensory predictions. The goal was to eventually conduct translaminar multi-electrode recordings to test whether neuronal activity patterns in the mouse visual cortex carry decodable predictive representations, and whether these are localised to any cortical layer or frequency band. I found that, while mice successfully acquired the first, stimulus location discrimination task, their behaviour showed no evidence of using auditory cues to produce orientation specific stimulus predictions. Furthermore, mice failed to achieve above-chance performance on the second, same/different orientation judgement task. Thus, neither task was appropriate for exploring predictive coding in the visual system of the behaving mouse.</p>
<p>In the second chapter, I conducted two series of experiments, combining local field potential recordings with optogenetics, to explore whether the proposed oscillatory signatures of predictive coding can be found in a local circuit of mouse primary somatosensory cortex (S1) in vitro. Slices of S1 expressing the photosensitive channelrhodopsin under the CaMKII promoter received stimulus patterns which followed particular statistical norms. On a minority of trials, norms were violated with stimuli which either deviated from the expected stimulation amplitude (Exp 1) or locus of stimulation within the barrel column (Exp 2). Both experiments provided no evidence that power or peak frequency in the γ-band - the alleged carrier of prediction error signals - was sensitive to stimulus (un)predictability. Furthermore, I obtained no evidence that power in the α-band represented sensory predictions, on an acute or long-term temporal basis. α-power did not differentiate between conditions with diverging likely futures, and failed to scale with increasing stimulation history which goes hand-in-hand with accrual of sensory expectation (van Ede et al., 2010; van Pelt et al., 2016).</p>
<p>In the third chapter, I combined patch-clamp electrophysiology and computational modelling to explore how single cell computation is affected by the γ rhythm - a critical information carrier in the predictive coding scheme. Previous research has shown that γ-modulated excitation boosts the input-output slopes of cortical pyramidal cells over and above what can be achieved with arrhythmic excitation (Sohal et al., 2009). However, it is still unclear whether the same enhanced response properties can be achieved with γ-rhythmic inhibition, which is critical for inducing network γ oscillations (Mann et al., 2005; Traub et al., 1996; Whittington et al., 1995), and whether the γ-rhythm is unique in this respect. This question was explored by examining spike patterns of putative pyramidal cells in mouse S1 in response to excitatory current pulses of increasing amplitudes, while simulated rhythmic inhibitory post-synaptic potentials (IPSGs) were administered with varying degrees of γ-power, using dynamic clamp. The experiment revealed that, with increasing γ-power in the IPSG, pyramidal cells achieved higher maximal spike rates, as well as exhibiting higher input-output slopes and spike timing fidelity. A further experiment was conducted to compare the effects of varying the frequency of inhibition, finding that γ-frequency inhibition (40 Hz) appeared to be optimal in its ability to boost the input-output properties of pyramidal cells. A computational model of the single cell was able to replicate these effects by simulating conditions of low M-type K+ channel conductance and high flux through voltage-gated Ca2+ channels, shedding some light on the potential underpinnings of γ’s unique effects on single-cell computation.</p>
|
|---|