top of page
Russell Poldrack
Incidential influences on value
Most work on the acquisition and modification of value representations has focused on reinforcement learning mechanisms. I will discuss a program of research that has investigated incidental influences on value, particularly through approach and avoidance experiences. This work shows that approaching a stimulus in a specific manner can have lasting effects on value representations. The effects of inhibitory avoidance on value representations are much less clear.
Michael Frank
Understanding statistical word learning
in a social context
How do children learn their first words? Even very young children make use of both distributional and social information. But it may not make sense to cast statistics and social interaction as two distinct cues or information sources. Instead, I will argue that much of children's early word learning is best explained via the assumption that learners infer speakers' communicative intentions in the presence of uncertainty about word meanings. These joint social/statistical inferences allow children to learn words both within and across situations, but they are constrained by children's developing memory and attention. I'll show data on the value of this communicative inference framework for thinking about phenomena like mutual exclusivity and Gricean pragmatics but also highlight developmental limitations, proposing a synthesis between computational- and algorithmic-level accounts of early word learning.
Floris de Lange
How does the past bias the present?
Sensory signals are highly structured in both space and time. These regularities allow expectations about future stimulation to be formed, thereby facilitating decisions about upcoming visual features and objects. One such regularity is that the world is generally stable over short time scales. Temporal context could be exploited by the brain in two opposite ways. On the one hand, biasing of current stimuli towards the past could be beneficial in terms of promoting visual stability (Fischer & Whitney 2014). On the other hand, repulsion of current stimuli away from the past can boost sensitivity of the visual system to changes in the physical environment (Muller et al 1999).
In my talk I will discuss recent psychophysical data from my lab that both processes occur: while perception is repelled away from the past, post-perceptual decision processes are attracted towards the past. These opposite biases on perceptual and post-perceptual processes may imbue the nervous system with an optimal balance between sensitivity and stability that is required to operate successfully in the environment.
In the second part of my talk, I will discuss how spatial context changes sensory computations in the primary visual cortex via feedback excitatory and inhibitory processes.
Ram Frost
Explaining visual statistical learning performance:
A perspective of Information theory
In order to extract the statistical structure underlying a continuous sensory input, the individual elements constituting the stream have to be encoded and their transitional probabilities (TPs) should be learnt. This suggests that variance in statistical learning (SL) performance reflects efficiency in encoding representations as well as efficiency in detecting their distributional properties (Frost et al., 2015). However, a recent study suggests that the encoding of visual shapes and the computation of their TPs are not independent processes, one preceding and feeding into the other. Rather, these two processes display substantial interaction: Sensitivity to extent of the TPs of elements in the stream is modulated by their exposure duration (ED), and susceptibility to ED of elements is modulated by the their predictability (Bogaerts et al., 2016). Here, we entertain the theoretical hypothesis that one unifying construct – the extent of information per time unit – can account for this counter-intuitive pattern of performance. This theoretical approach blurs the distinction between constraints related to encoding of events (e.g., ED) and constraints related to learning their regularities (e.g., TPs between elements), merging them into one processing principle. It also conforms with recent neurobiological evidence regarding differential sensitivity of neurons to levels of entropy. The theoretical and methodological implication of this approach will be discussed.
Zoe Kourtzi
Adaptive computations for flexible cognition
in the human brain
Successful interactions in complex environments entail making optimal decisions about the present and predictions about the future. Extracting the key features from our sensory experiences and deciding how to interpret them is a computationally challenging task that is far from understood. Accumulating evidence suggests that the brain may solve this challenge by combining sensory information and previous knowledge about the environment acquired through evolution, development, and everyday experience. We combine behavioral and brain imaging measurements to investigate the role of learning and experience-dependent plasticity in optimizing decisions. We demonstrate that learning translates sensory experiences to decisions about the present scene and predictions about upcoming events by shaping cortico-subcortical circuits. Our findings propose that long-term experience and short-term training interact to shape the optimization of decision and prediction processes in the human brain.
SCHEDULE
1:30pm // Registration
1:50pm // Welcome
2:00pm // James McClelland
3:00pm // Daphne Bavelier
4:00pm // Break
4:20pm // Nicholas Turk-Browne
5:20pm // Discussion 1
9:00am // Floris de Lange
10:00am // Marvin Chun
11:00am // Break
11:20am // Alexandre Pouget
12:20pm // Discussion 4
1:00pm // End
9:00am // Beverly Wright
10:00am // Ram Frost
11:00am // Break
11:20am // Zoe Kourtzi
12:20pm // Discussion 2
1:00pm // Lunch
2:00pm // Daphna Shohamy
3:00pm // Russell Poldrack
4:00pm // Break
4:20pm // Michael Frank
5:20pm // Discussion 3
7:30pm // Banquet dinner
FRIDAY, JAN 6th
SATURDAY, JAN 7th
SUNDAY, JAN 8th
Zoe Kourtzi
Ram Frost
Floris de Lange
Russ Poldrack
Mike Frank
Bev Wright
James McClelland
How expertise arises and how it facilitates new learning
Expertise, it has been said, is something that is acquired gradually with experience – to become a true expert might take 10 years or 10,000 hours of practice. We can also think of cognitive development in domains like reading, language processing, knowledge about natural kinds, and even mathematical cognition as the gradual accumulation of expertise. Yet a degree of expert-like performance can be exhibited over shorter practice periods, so that the emergence of expertise in humans and animals can be brought into study in the laboratory. What is it that is occurring as expertise is being acquired? How does expertise change the way we process and learn new things in the domain of our expertise? I will address these questions by broadly drawing on findings from research on expertise in action games and thought games like chess or go, and from research on ‘schema acquisition’ in animals and human cognitive development. Most of my suggestions for answers to the questions raised above will be couched within the framework of complementary learning systems theory (McClelland, McNaughton & O’Reilly, Psych. Rev., 1995; Kumaran, Hassabis & McClelland, TiCS, 2016) and related approaches to the development and neural basis of learning and memory.
Jay McClelland
Marvin Chun
Probing neural representations of learning and memory with fMRI repetition suppression, pattern similarity, and convolutional neural nets
As an elemental form of learning, stimulus repetition influences memory and corresponding neural activity, while fMRI reveals different neural signatures associated with such changes. Repetition suppression refers to reductions in neural activity with repetition (e.g., Buckner and Schacter, 1997), while multivariate pattern analysis measures the similarity of neural activity patterns across repetition (Xu et al., 2010). When we directly compared the two measures in a scene categorization and subsequent memory task, we found a double dissociation in which repetition suppression, but not pattern similarity, predicted repetition priming, while pattern similarity, but not repetition suppression, predicted explicit memory (Ward et al., 2013). To gain further insight into the neural code of learning, we are currently developing new methods based on hierarchical convolutional neural networks (HCNN) trained for visual recognition. Preliminary evidence suggests that visual category selectivity in the human brain may be supported by filter-based representations similar to those in HCNNs, and that learning alters category-selectivity within these representations.
Daphne Bavelier
Daphne Bavelier
Learning to learn: Lessons from Action Video Games
A vexing issue in the field of learning is that, while we understand how to promote superior performance through practice, the resulting behavioral enhancement rarely extends beyond the practiced task. Such learning specificity is a major limitation for effective interventions, whether educational or clinical ones. Here we will consider first how learning and generalization may be enhanced, through a mechanism we term ‘learning to learn’ (L2L). We then ask what may be the determinants of ‘learning to learn’ – differentiating between adjusting parameters as learning of a specific task proceeds from extracting the structure across tasks to facilitate learning and generalization.
Marvin Chun
Nicholas Turk-Browne
Multiple brain systems supporting learning and memory
Dogma states that memory can be divided into distinct types, based on whether conscious or not, one-shot or statistical, autobiographical or factual, sensory or motor, etc. These distinctions have been supported by dissociations in brain localization, task performance, developmental trajectories, and pharmacological interventions, among other techniques. A natural consequence is the assumption of a one-to-one mapping between brain systems and memory behaviors. Aside from theoretical concerns about dissociation logic, there have also now been several empirical demonstrations of where these boundaries break down, from contributions of the hippocampus to reward learning and motor behavior to rapid episodic-like learning in frontal cortex. These considerations suggest that behavior is overdetermined by multiple memory systems. As a case study, I will describe a series of neuroimaging, neuropsychological, and computational studies implicating the hippocampal system in statistical learning, a function more traditionally ascribed to cortical systems. I will end by considering some open questions that arise from this perspective, including about how the function and balance of memory systems changes over development and how multiple memory signals are integrated to guide behavior.
Nick Turk-Browne
Daphna Shohamy
Learning structure in uncertain environments
Adaptive learning requires updating of expectations based on changes in reinforcement. Extensive work has revealed the basic neural and cognitive mechanisms by which such reinforcement learning is accomplished, identifying prediction-error driven mechanisms in the striatum. However, for all the progress on this question, this work has focused almost entirely on a very rudimentary form of learning - the association of a single cue with a single outcome – and on the role of a single circuit. But almost any real world environment requires learning about more complex configurations of cues. Moreover, learning about the structure of such configurations and their association with outcomes requires integration of sensory, motor and motivational information across time. I will present behavioral, neural and computational findings that begin to address questions about how the value of configurations of cues are learned, represented and updated. Our findings suggest that such learning involves interactions between the hippocampus and the striatum, as well as a broader network of motor and sensory regions. This work points to an important role for the hippocampus in building relational representations that are central to adaptive learning.
Daphna Shohamy
Alexandre Pouget
Confidence, uncertainty and learning
Alex Pouget
Beverly Wright
Induction and prevention of auditory perceptual learning
Performance on many perceptual tasks improves with practice. My coworkers and I have been investigating the factors that induce and those that prevent perceptual learning on auditory skills. For example, we have evidence that (1) for improvement to occur across days requires a sufficient amount of training per day, within a restricted time period, and that additional daily training can be superfluous, (2) a combination of practice with relevant auditory stimuli and additional stimulus exposures without practice can enhance learning, and (3) practice on more than one auditory task within a single session can either promote or disrupt learning depending on how the training trials are distributed. Conclusions drawn from learning on fine-grained auditory discrimination tasks have held for visual and speech learning, suggesting that common principles are at play across multiple domains. Knowledge of these issues will lead to more effective perceptual training strategies to aid rehabilitation and promote skill enhancement.
bottom of page