Adaptive coding for dynamic sensory inference

Abstract

Behavior relies on the ability of sensory systems to infer properties of the environment from incoming stimuli. The accuracy of inference depends on the fidelity with which behaviorally relevant properties of stimuli are encoded in neural responses. High-fidelity encodings can be metabolically costly, but low-fidelity encodings can cause errors in inference. Here, we discuss general principles that underlie the tradeoff between encoding cost and inference error. We then derive adaptive encoding schemes that dynamically navigate this tradeoff. These optimal encodings tend to increase the fidelity of the neural representation following a change in the stimulus distribution, and reduce fidelity for stimuli that originate from a known distribution. We predict dynamical signatures of such encoding schemes and demonstrate how known phenomena, such as burst coding and firing rate adaptation, can be understood as hallmarks of optimal coding for accurate inference.

Efficient and adaptive sensory codes

“…At any point in time, the most discriminative features are determined by the observer's dynamically evolving belief about the stimulus distribution. The online estimation of changes in stimulus distributions from limited and noisy stimulus samples is thus an inherently complex problem, and is further complicated if the representation of these stimuli is lossy [80]. As a result, there will necessarily be periods of time when the system incorrectly estimates the stimulus distribution and the code is maladapted.…”

“…Because we used decoding error to evaluate performance, the optimal solution that we derived ultimately favors local discrimination, and uses detection only as a means of improving discrimination. Other systems might instead favor the detection of changes in context over the discrimination of local stimulus details [80].…”

“…Because the context can only take two values, the posterior can be summarized by the probability that the environment is in the high context at time t, given by P H t = p(c t = c H |ŝ τ ≤t ). On each timestep, P H t was updated with incoming stimulus estimates (as originally derived in [31] for raw stimulus values; also see [80]). We averaged the posterior to compute the point estimateĉ t = P H t c H + (1 − P H t )c L , which could then be used to estimate the stimulus distribution p(s t |ĉ t ) (where p(s|ĉ) = N (s;ĉ, 1) in a mean-switching environment, and p(s|ĉ) = N (s; 0,ĉ 2 ) in a variance-switching environment).…”

Representing utility and deploying the body

“…The form of the performance term differs widely across models. Some models commit to a specific task and an associated behavioral objective, such as estimation or tracking error (Mackowiak and Wiederholt 2009;Młynarski and Hermundstad 2018;Park and Pillow 2017;Sims 2003;Sims et al 2012; van den Berg and Ma 2018), categorization accuracy (Li et al 2017;Młynarski and Hermundstad 2018;van den Berg and Ma 2018), or discriminability (Ganguli and Simoncelli 2014). Other models instead use mutual information between stimulus and response as a performance term (Barlow 1961;Laughlin 1981;Olshausen and Field 1996;Wei and Stocker 2015;Zaslavsky et al 2018).…”

Opportunities for emotion and mental health research in the resource-rationality framework

“…Resource-rational models are often non-committal about the timescales over which the optimization occurs. Recent work on efficient codes in nonstationary environments (Młynarski and Hermundstad 2018) holds promise for bridging the divide.…”