Sources of predictive information in dynamical neural networks

Abstract

Behavior involves the ongoing interaction between an organism and its environment. One of the prevailing theories of adaptive behavior is that organisms are constantly making predictions about their future environmental stimuli. However, how they acquire that predictive information is still poorly understood. Two complementary mechanisms have been proposed: predictions are generated from an agent’s internal model of the world or predictions are extracted directly from the environmental stimulus. In this work, we demonstrate that predictive information, measured using bivariate mutual information, cannot distinguish between these two kinds of systems. Furthermore, we show that predictive information cannot distinguish between organisms that are adapted to their environments and random dynamical systems exposed to the same environment. To understand the role of predictive information in adaptive behavior, we need to be able to identify where it is generated. To do this, we decompose information transfer across the different components of the organism-environment system and track the flow of information in the system over time. To validate the proposed framework, we examined it on a set of computational models of idealized agent-environment systems. Analysis of the systems revealed three key insights. First, predictive information, when sourced from the environment, can be reflected in any agent irrespective of its ability to perform a task. Second, predictive information, when sourced from the nervous system, requires special dynamics acquired during the process of adapting to the environment. Third, the magnitude of predictive information in a system can be different for the same task if the environmental structure changes.

Predictive coding is a consequence of energy efficiency in recurrent neural networks

“…This tight link has been established before. 31 , 32 , 33 , 34 , 35 , 36 In our work, the link can be explained in the following way: the preactivation allows “predictive inhibition” while forcing the network weights to be as small possible so that irreducible noise has as little impact as possible on the network activity. We refer the reader to the supplemental information for a derivation of why this is the case.…”

“… 34 demonstrated that a system with memory exposed to a stochastic signal must be predictive to operate at maximal energy efficiency. Candadai and Izquierdo 35 showed, information theoretically, that predictable environments produce neural networks that exhibit predictive information. In the work by Sengupta et al., 60 the minimization of thermodynamic energy was linked to information processing and efficiency using the Jarzynski equality.…”

“…In search for such principles, we take inspiration from the organism’s limited energy budget and previous established links between prediction and energy efficiency. 31 , 32 , 33 , 34 , 35 , 36 Expanding previous work on efficient coding and sparse coding, 37 , 38 , 39 , 40 , 41 , 42 , 43 we subject recurrent neural networks (rate-based recurrent neural networks [RNNs]) to predictable sequences of visual input and optimize their synaptic weights to minimize the largest sources of energy consumption in the mammalian cortex: action potential generation and synaptic transmission. 44 , 45 , 46 , 47 , 48 , 49 We then test the resulting networks for phenomena typically associated with predictive coding.…”

Integrated World Modeling Theory (IWMT) Expanded: Implications for Theories of Consciousness and Artificial Intelligence

“…NSCs consist of recurrent neural networks (RNNs) organized into a predictive hierarchy, where each RNN attempts to predict data from the level below, and where only unpredicted information is passed upwards. These predictions are realized in the form of recurrent dynamics (Candadai & Izquierdo, 2020;Lu & Bassett, 2020), whose unfolding entails a kind of search processes via competitive/cooperative attractor formation over states capable of most efficiently responding to-or resonating with (Rumelhart & McClelland, 1987;Safron, 2020b)ascending data streams. In this way, much as in FEP-AI, predictive abilities come nearly "for free" via Hamilton's principle of least action (Ali et al, 2021;K.…”

Conservative Significance Testing of Tripartite Interactions in Multivariate Neural Data

“…For example, PID has been used to demonstrate a relationship between synergy and feed-back information flow in mouse organotypic cultures [69], to show significant synergy between somatic and apical dendritic output of L5b pyramidal neurons and its relationship to activation of dendritic GABA B receptors in rat S1 slices [68], and to estimate unique contributions of acoustic features of speech to BOLD responses in humans [22, 23]. Further, it has been used to explore the structure of simulated input-driven recurrent network models [16] and artificial generative neuronal networks [75]. We believe that PID will be increasingly applied in coming years, especially in studies addressing non-linear confounding effects, the specificity of functional relations, and synergistic encoding.…”