Conference on Cognitive Computational Neuroscience (CCN)
Understanding how complex cognitive processes are carried out by the brain is a problem that is bigger than any one discipline. Historically, distinct disciplines have attacked different aspects of the problem, but often in isolation. Each has it’s own strengths and weaknesses. For example, at one end, cognitive neuroscience has provided models that can explain measured neural activity, but are difficult to scale to understand more complex higher-level cognitive behavior. At the other end, cognitive science has developed general computational principles behind human cognition, but are difficult to relate the underlying neural physiology. Machine learning and artificial intelligence have made great strides in engineering artificial system to autonomously solve complex cognitive tasks, but these are difficult to relate to the human brain.
We believe that the time is ripe for a convergence of cognitive neuroscience, cognitive science, and AI. This would ideally result in a new discipline that addresses the questions of cognitive neuroscience by merging its experimental techniques with the powerful concepts and theories that have emerged in cognitive science and AI.
Currently, there are institutional and historical barriers to merging these three disciplines. Researchers in cognitive neuroscience, cognitive science, and AI generally take appointments in different academic departments, attend different conferences, and publish in different journals. Meetings in cognitive science and neuroscience are often atomized, focusing on a specific perceptual modality or cognitive skill. Meanwhile, AI conferences are dominated by engineering-oriented presentations of system architecture and performance, with little reference to neurobiology. As a result, opportunities for researchers in these three disciplines to interact and exchange ideas is far too limited.
We therefore propose Cognitive Computational Neuroscience (CCN), an annual scientific meeting for researchers interested in characterizing the neural computations that underlie complex human behavior. CCN will be organized around a shared commitment to to developing models of brain information processing that are (1) fully computationally defined and implemented in computer simulations, that are (2) neurobiologically plausible, that explain rich measurements of (3) brain activity and (4) behavior for naturalistic stimuli and tasks, and that (5) perform feats of intelligence such as recognition, internal modeling and memory of the environment, language production, decision-making, planning, action, and motor control.
CCN will encourage participation from all experimentalists and theoreticians investigating complex cognitive abilities from a computational point of view. We expect that CCN will draw researchers that address the following challenges:
Understanding brain information processing underlying real-world tasks that involve natural stimuli, rich knowledge, complex inferences, and behavior
Revealing principles of brain connectivity and dynamics at multiple scales
Developing cognitive- or neural-level models of perception, cognition, emotion, and action
Using brain and behavioral data to test these models
Understanding commonalities and differences between neurobiological and artificial systems
Using techniques from machine learning and artificial intelligence to model brain information processing, and, conversely, incorporating neurobiological principles in machine learning and artificial intelligence
Measuring brain activity at multiple spatial scales in humans and animal models
Using psychophysical techniques to relate sensory inputs to behavioral responses
The CCN conference will thus seek to foster collaboration between these three key disciplines – computational neuroscience, cognitive science and AI – see the Figure below.
Cognitive neuroscience has approached the problem by measuring and studying the patterns of brain activity that occur during the performance of cognitive tasks. Its major strength has been pioneering the development of non-invasive imaging techniques that can be applied directly to humans, and dense array recording of neurons in primates trained to perform complex tasks. These datasets have the potential to reveal how cognition is implemented in the brain; however, making sense of these rich patterns of brain-wide activity has been a tremendous challenge. Researchers have expended correspondingly tremendous effort on developing statistical methods for observing patterns in and testing hypotheses against their data [#REFS]. Perhaps as a side-effect of this focus on method development, many of the models of cognition that are tested in cognitive neuroscience experiments are of the box-and-arrow variety [#REF], and do not reflect the sophistication of current computational models of cognitive processing.
Cognitive science has addressed the problem by developing fully defined computational models of cognition that account for patterns in human behavior. It’s strength has traditionally been in identifying models of computation that double as plausible models of cognition. Two notable examples are connectionism (which seeded the current revolution in deep learning, see below) [#REFS] and Bayesian or probabilistic models of cognition [#REFS]. A weakness of the field is that these models often fail when scaled up to natural, real-world complex tasks, relegating much of cognitive science to the study of instructive but unrealistic toy problems.
Artificial intelligence (AI) approaches cognition by developing systems that autonomously perform specific cognitive tasks. It has emerged as an extremely potent field of research: recent years have seen the development of systems that solve cognitive tasks that are as or more complex than those isolated is typical neuroscience experiments [#REFS]. Interestingly, many of the most successful AI systems are deep neural networks (DNN) [#REF], which are directly inspired by neurobiology. The weakness of AI is that it is an essentially engineering discipline; although DNNs have achieved human-level performance using neural-like architectures, the job of understanding how (and if) these systems relate to the brain has only just begun [#REF]. It is also important to note that, although impressive, AI systems still trail far behind humans at solving complex tasks. It is likely that the gap will be closed only by importing new ideas and concepts from cognitive science and neuroscience into AI (which is in fact how the field began).
CCN will be unique in its focus on the intersection of disciplines in which comprehensive theories of complex brain computations involving rich world knowledge are emerging. CCN will contribute to a stable, virtuous cycle of interaction between disciplines, in which neurobiological experiments inspire the discovery of new computational models, and new computational models inspire the design of neurobiological experiments.
The format of the conference will have several unique features that will facilitate interaction across disciplines:
Hands on tutorials: While many conferences have had tutorial sessions on state-of-the-art research, CCN will include more fundamental educational sessions to introduce researchers from outside the discipline. Importantly, preference will be given to tutorial sessions that include “hands on” components that enable students to work on real data to maximize the value of their work.
Debate and cage matches: To encourage debate, CCN will also include “cage matches” where two or more prominent researchers in the field will argue opposite positions in key neuro-scientific problems. The debate will be moderated by “socratic” examiner that will push both sides to refine their arguments.
Mix of poster and oral presentations: Following highly successful formats for machine learning conferences, the conference will include a mix of poster sessions, where researchers can spend significant time with individual posters of interest; and oral talks that are of wide importance to the entire field.
We envision CCN not only as a engine for advancing research, but as vehicle for making broader impacts on education and society. In its early stages, the educational focus of CCN will be on increasing the visibility of women scientists in computational fields. See below for details on our proposed efforts.