↑ Return to Research

Interdisciplinary work on the boundary between Psychology and Computer Science

Our research on Cognitive Neuroscience is frequently aimed at understanding mechanisms by which the human brain accomplishes cognition. Our hypotheses are frequently inspired by and driven by analogy to technology and computer science, both when there is commonality with human intelligence and also where there are distinct differences.

For example, I was asked a few years back if I could answer a question about the capacity of human memory for Scientific American Mind. To answer, two points of contact were drawn: digital video recording and computational constraints on human memory. First, what if we had a DVR running in our brains recording our lives for every waking moment? How much storage space would we need? This is a pretty big number but it’s calculable (~1e15 bytes). Second, how much storage can we hold in a large population of neurons (e.g., 1 billion, 1e9). If storage is linear with the number of neurons, we are going to run out of space far too quickly. However, it appears that the brain uses distributed codes across neurons which allows for storage capacity that is exponential with the number of neurons. That provides easily ample storage space, but highlights a secondary problem. It is very difficult to store a precise memory trace quickly in a distributed system due to a problem described as catastrophic interference in known models of connectionist networks. A key computational approach is to use a two-step system with an initial step of pattern separation in a smaller capacity system from which memories are then consolidated into a high-capacity distributed system. Unfortunately, this neurobiological process in the brain is slower than we actually experience the world. The reason we don’t remember everything is not that we run out of long-term storage space, but that the bandwidth of information transfer into long-term memory does not keep up with our experiences. So the answer to the original question of “how much memory?” is essentially: a lot, enough for a human lifetime (unless we start living a lot longer or find a way to get our brains to store memories at a greatly increased rate), but practically that still doesn’t mean we will remember everything we experience.


In our laboratory studies of implicit learning, we examine paradigms in which we induce learning without awareness. That is, participants improve task performance but cannot consciously explain why they have improved after practice. The ability to do something without being able to explain how is often associated with highly expert performance. However, we have shown that this kind of learning can occur quickly in specialized paradigms that we use (and create) in our lab. We observe robust implicit learning in 10-20 minutes of practice suggesting that this kind of learning is likely part of how we learn about the world and does not just emerge after many hours of dedicated practice.

In our recent perceptual-motor sequence learning work, we realized that our lab tasks were essentially storing an arbitrary sequence of information in a participant’s brain that could potentially act as authenticating information like a password. Through a collaboration with Pat Lincoln (computer scientist, SRI) and Dan Boneh (cryptographer and computer scientist, Stanford) we demonstrated that this paradigm can provide a solution to a cryptographic weakness known as a “rubber hose” or coercion attack. The attack is based simply on forcing the user to provide their password to the attacker so that the attacker can impersonate the victim. An implicitly learned password, however, cannot be verbally reported and thus cannot be coerced out of a user.

In our approach, the user goes through our learning task which is designed to look like a video game (a rhythm game like Guitar Hero or Dance, Dance, Revolution). The motor patterns they produce are not tied to auditory information (songs) but are constructed mathematically. The user’s brain extracts the repeating structure of training and then they can show us that they know the sequence by improved performance compared to untrained sequences. However, when shown the sequence, the user cannot recognize it, nor can they report it verbally. That is, their “fingers” know the sequence but the parts of the brain that support conscious memory do not. In this state, they can authenticate by showing they know their sequence, but could never tell anybody else what the sequence is, even under coercion.

With support from an NSF grant on innovative approaches to cybersecurity, we showed that we can entrain a cryptographically strong individualized sequence (~30 bits in a covertly learned 30-item sequence) and identify individuals fairly well. Some logistical challenges remain to practical use, such as the time needed to learn and express the password knowledge. Ongoing research aims to improve on these via a variety of approaches such as machine learning to incorporate individual differences in motor performance and electroencephelography to identify neural signals of knowledge.

Computational models of learning

To capitalize on these kinds of applications of implicit learning, our lab works to improve our basic understanding of how the brain extracts this kind of structure from experience.  We can calculate the complexity of the statistical structure in the information we entrain, but we do not yet know the computational parameters of the brain’s implicit learning system. We have not yet identified any bandwidth constraints such as exist for explicit memories, but the relatively simpler neural circuits supporting implicit memory a very likely to be capacity constrained. That is, it seems like we must run out of space, or that these brain circuits must use some short-cuts in representation to do things like rapidly learn very long and complex sequences.

Our approach to this question is based on building simulation models based on different hypotheses about what kinds of mathematical structure the brain might learn. For example, an ongoing project is comparing a mathematical model of the simplest possible statistics that could be learning in our basic experimental paradigm. We use sequences that could theoretically be learned simply as fragments of three successive responses by the most efficient mechanism. However, this does not appear to be what the brain does. Early evidence is that these simple neural circuits are able to extract statistical information across fragments of five or six responses almost immediately during learning. Thus, using this simulation modeling approach, we find evidence that these neural circuits are likely to be much smarter than we originally anticipated.

Intuition: Systems level modeling

We have also applied simulation modeling to systems-level hypotheses related to the use of both conscious and nonconscious memory in decision making. The implicit learning we study in the lab can influence performance directly (e.g., as in motor sequences) but it also frequently produces an intuitive hunch about a decision. In our laboratory experiments, we generally have to encourage participants to guess or to go with a gut instinct to get them to make responses based on implicit memory. In these cases, we see the influence of implicit learning on intuition-like guesses that are better than chance, indicating that learning occurred prior to the guess.

One of the challenges embedded in these paradigms is getting participants to make responses based on this intuitive sense. Participants often have some explicit knowledge, but we have designed the tasks to make this type of memory unhelpful. In some cases, however, participants can rely too much on the explicit memory and this blocks their ability to use their implicit knowledge.

We have developed a computational model, PINNACLE, to capture this kind of competitive interaction between conscious and nonconscious memory. Using functional neuroimaging, we have shown that we can separate the neural signatures of implicit and explicit knowledge and guided by the model, even identify executive function processes associated with the decision to rely on one’s intuition. This work is the basis of a recently funded project from the Office of Naval Research aimed at improving training of decision making in marines, especially squad leaders. Using our basic science, driven by the PINNACLE model, of multiple memory systems and decision-making, we will develop candidate training improvements that that will lead to more rapid development of accurate intuitive decision making sense for military personnel.


Brain training and gamification

Simulation-based training frequently uses virtual reality environments to provide the opportunity for skill practice repetitions inexpensively. These environments frequently look like video games, often bringing in ideas like “gamification” to improve training. Video games themselves often involve significant skill learning that is acquired over many hours of practice within the video game environment. We hypothesize that in all of these cases, many of the performance improvements depend on the same neural changes that reshape brain systems during implicit learning tasks.

From this perspective, brain training that occurs in any game or simulation environment should be strongly influenced by implicit learning mechanisms of the kind we capture in our laboratory research. A crucial question becomes that of transfer, the ability to apply this implicitly acquired information in novel contexts. Our own work shows that implicit learning is inflexible, posing constraints on transfer. Yet we also know that the transfer problem can be overcome since experts rapidly acquire new related skills in their area of expertise. In studies of protocols aimed at creating brain training (e.g., training working memory to improve fluid intelligence) and examination of surreptitious brain training from video games (improvement of visuo-spatial skills), there appear to be small but promising and robust transfer effects. The fact that the effects are small, which has led to some reports of failures to replicate, probably reflect the fact that the training protocols have not been optimized based on the underlying implicit learning mechanisms. The real potential of these approaches will not be realized until we have a better understanding of implicit learning mechanisms and better models of the training tasks and transfer domains they are applied to.

Leave a Reply