Dan YaminsRecent advances in computer vision and AI have made it possible to build deep neural networks that mimic aspects of computation in the primate and human visual system. The core idea behind these results is task-driven modeling, e.g. optimize a neural network for a complex ecologically relevant behavior, and then compare the learned model to neural responses in the brain areas thought to underlie that behavior. While this approach has produced powerful predictive models of neural representations in adult animals, its main successes so far have unfortunately relied on using heavy semantic supervision, using large labeled datasets - data streams to which real animals do not have access. I will discuss recent work my lab has been doing to move beyond this limitation, developing less heavily supervised approaches to train deep neural network models of vision that may be more plausible models of real neural learning. These ideas rely on loss functions defined in interactive worlds that attempt to better capture the true complexities of the environment in which early juvenile development takes place.
Matt BotvinickTwo decades of neuroscience research on reward-based learning has converged on a canonical model, under which the neurotransmitter dopamine 'stamps in' associations between situations, actions and rewards by modulating the strength of synaptic connections between neurons. However, a growing number of recent findings have placed this standard model under strain. This talk draws on recent advances in AI to introduce a new theory of reward-based learning. Here, dopamine trains another part of the brain, the prefrontal cortex, to operate as its own freestanding learning system. This perspective accommodates the findings that motivated the standard model, but also deals gracefully with a wider range of observations, laying a fresh foundation for next-generation research.
Joshua Tenenbaum In its most general form, perception can be defined as the solution to an inverse problem: identifying the world scene that gave rise to the observed retinal (or auditory, or haptic) data. It remains a mystery how the brain constructs such rich representations of the geometry of objects in a scene and their physical properties extremely quickly, in at most a few hundred milliseconds. Traditional models of how the brain solves such inference problems use iterative methods that are hard to map onto neural circuits and much too slow to explain online perception. Here we present a new approach that combines a deep neural network with a probabilistic generative model. This approach is as fast as pure feedforward models, but generates a rich description of 3D object shapes and physical properties. Applied to faces, our model explains both the tuning properties of cells in the macaque face patch system and human behavior in recognizing familiar and unfamiliar individuals.
Joshua PetersonRecent neural network models have resulted in significant progress in computer vision, speech recognition, and natural language processing, by learning representations of the statistical structure of complex visual, auditory, and linguistic stimuli. Understanding how the resulting models work -- and how well they correspond with human perception -- is an interesting scientific challenge. However, the representations that these models discover, when treated just as representations of complex stimuli, also offer the opportunity to extend the scope of psychological research. Psychologists studying problems such as categorization or memory have tended to focus on very simple stimuli that can be carefully controlled and parameterized. This makes it possible to formulate precise theories, but at the cost of potentially losing sight of the original phenomena: Do the same models that predict how people categorize sinusoidal gratings explain how they differentiate cats and dogs? I will talk about recent work that tries to leverage representations produced by neural networks -- for both images and language -- to study human cognition, highlighting the promise of this approach as well as some of the challenges.
Jessica HamrickMany machine learning systems are built to solve the hardest examples of a particular task, which often makes them large and expensive to run--- especially with respect to the easier examples, which might require much less computation. For an agent with a limited computational budget, this "one-size-fits-all" approach may result in the agent wasting valuable computation on easy examples, while not spending enough on hard examples. Rather than learning a single, fixed policy for solving all instances of a task, we introduce a "metacontroller" inspired by human cognition which learns to optimize a sequence of imagined internal simulations over predictive models of the world (called "experts") in order to construct a more informed, and more economical, solution. Our approach learns to adapt the amount of computation it performs to the difficulty of the task, as well as which experts to consult by factoring in both their reliability and individual computational resource costs. The metacontroller achieves a lower overall cost (task loss plus computational cost) than more traditional fixed policy approaches, demonstrating that our approach is a powerful framework for using rich forward models for efficient model-based reinforcement learning.
Tal LinzenLarge-scale artificial neural networks have shown great promise in natural language processing, reportedly reaching human-level performance in some tasks. Yet our understanding of the capabilities of these methods is typically limited to general statistics averaged across a random sample of texts. Such coarse-grained evaluation metrics stand in marked contrast to the rich array of highly specific patterns identified by linguists and cognitive scientists. I will argue that this detailed characterization of human-level knowledge of language provides a yardstick for the desired behavior of an artificial intelligence system. Applying it to neural networks can help us understand the strengths and weaknesses of existing architectures and analyze models that are otherwise difficult to interpret. Finally, neural networks that combine powerful statistical learning with different degrees of representational assumptions can serve as useful baselines for psychological modeling. I'll discuss case studies illustrating this approach in both syntax and semantics.
Noah GoodmanNoah Goodman will be participating in the panel discussion.