Toward an Integration of Deep Learning and Neuroscience

Abstract

Neuroscience has focused on the detailed implementation of computation, studying neural codes, dynamics and circuits. In machine learning, however, artificial neural networks tend to eschew precisely designed codes, dynamics or circuits in favor of brute force optimization of a cost function, often using simple and relatively uniform initial architectures. Two recent developments have emerged within machine learning that create an opportunity to connect these seemingly divergent perspectives. First, structured architectures are used, including dedicated systems for attention, recursion and various forms of short- and long-term memory storage. Second, cost functions and training procedures have become more complex and are varied across layers and over time. Here we think about the brain in terms of these ideas. We hypothesize that (1) the brain optimizes cost functions, (2) the cost functions are diverse and differ across brain locations and over development, and (3) optimization operates within a pre-structured architecture matched to the computational problems posed by behavior. In support of these hypotheses, we argue that a range of implementations of credit assignment through multiple layers of neurons are compatible with our current knowledge of neural circuitry, and that the brain's specialized systems can be interpreted as enabling efficient optimization for specific problem classes. Such a heterogeneously optimized system, enabled by a series of interacting cost functions, serves to make learning data-efficient and precisely targeted to the needs of the organism. We suggest directions by which neuroscience could seek to refine and test these hypotheses.

Visual pathways from the perspective of cost functions and multi-task deep neural networks

“…Common practice in machine learning is to express such an objective as a cost function (Domingos, 2012). As Marblestone and colleagues argue, the human brain can be thought of implementing something very similar to cost functions to quantify the collective performance of neurons and consequently to steer the learning of representations in a direction that improves a global outcome (Marblestone et al, 2016).…”

“…It has been argued that the brain can recruit local populations of neurons to assign local cost functions that enable fast updating of these neurons (Marblestone et al, 2016). Fig.…”

“…Such joint cost functions can be learned with a single network, where the degree to which shared representations (in the form of shared learned features) help or hurt with the optimization task is variable (Caruana, 1998). The shape of the cost has likely evolved such that they help make most sense of our environment (Marblestone et al, 2016): a loss may measure the absolute deviation from some target value, or the square of this difference, or any other mapping.…”

“…It was recently suggested that the brain uses a variety of cost functions for learning (Marblestone, Wayne, & Kording, 2016). These cost functions can be highly diverse.…”

“…We start with a discussion of the relevance of DNNs (LeCun, Bengio, & Hinton, 2015;Schmidhuber, 2015) and, following Marblestone (Marblestone et al, 2016), of cost functions for understanding the brain in Section 2. We extend our discussion with the importance of optimizing different cost functions simultaneously, presenting a hypothesis on the relationship between relatedness of tasks and the degree of feature representation sharing.…”