The target article by Lake et al. beautifully highlights limitations of today's artificial intelligence (AI) systems relative to the performance of human children and adults. Humans demonstrate uptake and generalization of concepts in the domains of intuitive physics and psychology, decompose the world into reusable parts, transfer knowledge across domains, and reason using models of the world. As Lake et al. emphasize, and as is a mathematical necessity (Ho & Pepyne Reference Ho and Pepyne2002), humans are not generic, universal learning systems: they possess inductive biases that constrain and guide learning for species-typical tasks.
However, the target article's characterization of these inductive biases largely overlooks how they may arise in the brain and how they could be engineered into artificial systems. Their particular choice of inductive biases, though supported by psychological research (see Blumberg [Reference Blumberg2005] for a critique), is in some ways arbitrary or idiosyncratic: It is unclear whether these capabilities are the key ones that enable human cognition, unclear whether these inductive biases correspond to separable “modules” in any sense, and, most importantly, unclear how these inductive biases could actually be built. For example, the cognitive level of description employed by Lake et al. gives little insight into whether the systems underlying intuitive psychology and physics comprise overlapping mechanisms. An alternative and plausible view holds that both systems may derive from an underlying ability to make sensory predictions, conditioned on the effects of actions, which could be bootstrapped through, for example, motor learning. With present methods and knowledge, it is anybody's guess which of these possibilities holds true: an additional source of constraint and inspiration seems needed.
Lake et al. seem to view circuit and systems neuroscience as unable to provide strong constraints on the brain's available computational mechanisms – perhaps in the same way that transistors place few meaningful constraints on the algorithms that may run on a laptop. However, the brain is not just a hardware level on which software runs. Every inductive bias is a part of the genetic and developmental makeup of the brain. Indeed, whereas neuroscience has not yet produced a sufficiently well-established computational description to decode the brain's inductive biases, we believe that this will change soon. In particular, neuroscience may be getting close to establishing a more direct correspondence between neural circuitry and the optimization algorithms and structured architectures used in deep learning. For example, many inductive biases may be implemented through the precise choice of cost functions used in the optimization of the connectivity of a neuronal network. But to identify which cost function is actually being optimized in a cortical circuit, we must first know how the circuit performs optimization. Recent work is starting to shed light on this question (Guergiuev et al. Reference Guergiuev, Lillicrap and Richards2016), and to do so, it has been forced to look deeply not only at neural circuits, but also even at how learning is implemented at the subcellular level. Similar opportunities hold for crossing thresholds in our understanding of the neural basis of other key components of machine learning agents, such as structured information routing, memory access, attention, hierarchical control, and decision making.
We argue that the study of evolutionarily conserved neural structures will provide a means to identify the brain's true, fundamental inductive biases and how they actually arise. Specifically, we propose that optimization, architectural constraints, and “bootstrapped cost functions” might be the basis for the development of complex behavior (Marblestone et al. Reference Marblestone, Wayne and Kording2016). There are many potential mechanisms for gradient-based optimization in cortical circuits, and many ways in which the interaction of such mechanisms with multiple other systems could underlie diverse forms of structured learning like those hypothesized in Lake et al. Fundamental neural structures are likely tweaked and re-used to underpin different kinds of inductive biases across animal species, including humans. Within the lifetime of an animal, a developmentally orchestrated sequence of experience-dependent cost functions may provide not just a list of inductive biases, but a procedure for sequentially unfolding inductive biases within brain systems to produce a fully functional organism.
A goal for both AI and neuroscience should be to advance both fields to the point where they can have a useful conversation about the specifics. To do this, we need not only to build more human-like inductive biases into our machine learning systems, but also to understand the architectural primitives that are employed by the brain to set up these biases. This has not yet been possible because of the fragmentation and incompleteness of our neuroscience knowledge. For neuroscience to ask questions that directly inform the computational architecture, it must first cross more basic thresholds in understanding. To build a bridge with the intellectual frameworks used in machine learning, it must establish the neural underpinnings of optimization, cost functions, memory access, and information routing. Once such thresholds are crossed, we will be in a position – through a joint effort of neuroscience, cognitive science, and AI – to identify the brain's actual inductive biases and how they integrate into a single developing system.
The target article by Lake et al. beautifully highlights limitations of today's artificial intelligence (AI) systems relative to the performance of human children and adults. Humans demonstrate uptake and generalization of concepts in the domains of intuitive physics and psychology, decompose the world into reusable parts, transfer knowledge across domains, and reason using models of the world. As Lake et al. emphasize, and as is a mathematical necessity (Ho & Pepyne Reference Ho and Pepyne2002), humans are not generic, universal learning systems: they possess inductive biases that constrain and guide learning for species-typical tasks.
However, the target article's characterization of these inductive biases largely overlooks how they may arise in the brain and how they could be engineered into artificial systems. Their particular choice of inductive biases, though supported by psychological research (see Blumberg [Reference Blumberg2005] for a critique), is in some ways arbitrary or idiosyncratic: It is unclear whether these capabilities are the key ones that enable human cognition, unclear whether these inductive biases correspond to separable “modules” in any sense, and, most importantly, unclear how these inductive biases could actually be built. For example, the cognitive level of description employed by Lake et al. gives little insight into whether the systems underlying intuitive psychology and physics comprise overlapping mechanisms. An alternative and plausible view holds that both systems may derive from an underlying ability to make sensory predictions, conditioned on the effects of actions, which could be bootstrapped through, for example, motor learning. With present methods and knowledge, it is anybody's guess which of these possibilities holds true: an additional source of constraint and inspiration seems needed.
Lake et al. seem to view circuit and systems neuroscience as unable to provide strong constraints on the brain's available computational mechanisms – perhaps in the same way that transistors place few meaningful constraints on the algorithms that may run on a laptop. However, the brain is not just a hardware level on which software runs. Every inductive bias is a part of the genetic and developmental makeup of the brain. Indeed, whereas neuroscience has not yet produced a sufficiently well-established computational description to decode the brain's inductive biases, we believe that this will change soon. In particular, neuroscience may be getting close to establishing a more direct correspondence between neural circuitry and the optimization algorithms and structured architectures used in deep learning. For example, many inductive biases may be implemented through the precise choice of cost functions used in the optimization of the connectivity of a neuronal network. But to identify which cost function is actually being optimized in a cortical circuit, we must first know how the circuit performs optimization. Recent work is starting to shed light on this question (Guergiuev et al. Reference Guergiuev, Lillicrap and Richards2016), and to do so, it has been forced to look deeply not only at neural circuits, but also even at how learning is implemented at the subcellular level. Similar opportunities hold for crossing thresholds in our understanding of the neural basis of other key components of machine learning agents, such as structured information routing, memory access, attention, hierarchical control, and decision making.
We argue that the study of evolutionarily conserved neural structures will provide a means to identify the brain's true, fundamental inductive biases and how they actually arise. Specifically, we propose that optimization, architectural constraints, and “bootstrapped cost functions” might be the basis for the development of complex behavior (Marblestone et al. Reference Marblestone, Wayne and Kording2016). There are many potential mechanisms for gradient-based optimization in cortical circuits, and many ways in which the interaction of such mechanisms with multiple other systems could underlie diverse forms of structured learning like those hypothesized in Lake et al. Fundamental neural structures are likely tweaked and re-used to underpin different kinds of inductive biases across animal species, including humans. Within the lifetime of an animal, a developmentally orchestrated sequence of experience-dependent cost functions may provide not just a list of inductive biases, but a procedure for sequentially unfolding inductive biases within brain systems to produce a fully functional organism.
A goal for both AI and neuroscience should be to advance both fields to the point where they can have a useful conversation about the specifics. To do this, we need not only to build more human-like inductive biases into our machine learning systems, but also to understand the architectural primitives that are employed by the brain to set up these biases. This has not yet been possible because of the fragmentation and incompleteness of our neuroscience knowledge. For neuroscience to ask questions that directly inform the computational architecture, it must first cross more basic thresholds in understanding. To build a bridge with the intellectual frameworks used in machine learning, it must establish the neural underpinnings of optimization, cost functions, memory access, and information routing. Once such thresholds are crossed, we will be in a position – through a joint effort of neuroscience, cognitive science, and AI – to identify the brain's actual inductive biases and how they integrate into a single developing system.