Distributed neural processing of self-generated visual input in a vertebrate brain

see the Fish Matrix

Direct brain state modulation by closed-loop feedback

Neurofeedback techniques are rapidly becoming established as effective treatment for various neuropathologies. In traditional neurofeedback protocols participants are shown a representation of their current brain activity guided by which they subsequently attempt to influence various frequency components. Training patients to suppress alpha frequency power has been shown to effective for treatment of ADHD [1]  and similar techniques have been used to  improve concentration, [2]. In this project we will explore an alternate protocol where we construct neurofeedback directly from  brain activity, circumventing a participants active involvement. Practically we will explore the feasibility that closed-loop interaction between high luminance LED and brain activity reported by EEG can arbitrarily modulate alpha frequency power across the occipital cortex.  This project will be a  collaboration with the Center for Computational Neuroscience and Robotics and Sackler Center for Consciousness science.

Recently theoretical and experimental work by Dr Buckley of the Centre for Computational Neuroscience and Robotics (CCNR)  has demonstrated  in primates the  feasibility of modulation brian dynamics by implementing real-time closed-loop visual feedback [3]. The goal of this project will be to investigate the feasibility of this human participants utilizing EEG.

It is already well  established  that exposing human participants to 10Hz high luminance visual flashes increase EEG alpha frequency (~10Hz) in visual areas [4]. Our  protocol represents  a novel modification whereby the sequence of visual flashing is derived directly from real-time EEG recordings of the participants brian. To implement we will first construct a pair of  visual stimuli goggles,  which contain high luminance (30000 lumens)  LED’s in each lense holder. We will then import  real-time EEG data into Matlab and use the same software to control the sequence of visual flashes. The participant will then be asked to where these goggle and close their eyes after which we we will deliver various flashing visual stimuli. Each EEG experiment will involve three different conditions. .

  • Standard flicker: Where participants are shown 10hz flashing stimulus. This will be used to validate our basic experimental setup.
  • Phase Reversal:  A 10 Hz flashing signal is presented but the phases of the signal will be intermittently shifted by half a cycle. This will attempt to validate the idea that brain activity is dependent on the moment-to-moment activity of the visual stimulus.
  • Flicker Feedback Suppression: The  visual  flashes are triggered when the raw EEG signal goes below a certain negative threshold.
  • Flicker Feedback Enhancement:  The  visual  flashes are triggered when the raw EEG signal goes above a certain positive threshold,

This project will involve a significant technological component involving the development of both novel hardware and software. It will also involve conducting human  EEG trials over summer under the supervision of Dr  Schwartzman. The significant research milestone are enumerated below.

Hardware and Software

  1. Develop protocol for streaming EEG data into software in real-time, this shoudl be possible using existing EEG equipment i and an off-the shelf Matlab plugin called Field Trip []. We will attempt reducing latencies wherever we can.
  2. Constructing the high luminance LED goggles set with 30000 lumens white LEDS and cables.
  3. Interfacing Matlab with the LED goggles via  National Instrument  Labjack board enabling us to digitally manipulate the LED flicker from the software.
  4. Using the real-time EEG stream to modulate the output flashing sequences of the LED’s

Free Energy Robots: A Bayesian route to autonomy in complex and dynamic environments.

Summary of Project

The idea that the brain is a prediction machine has rapidly begun to dominate the cognitive sciences. On this theory perception is a process of inferring the worldly causes of sensory data by minimizing error between actual sensations and those predicted by an internal probabilistic model. Furthermore, the “Free energy principle” (FEP) , a strong current formalism for the predictive brain hypothesis, allows for an account of action generation within the same framework. Specifically action is drawn as the process of modifying the world such that the consequent sensory input meets expectations encoded in the same internal probabilistic model. These two processes of inference, inferring the world and inferring actions needed to meet expectations, close the sensory/motor loop, and suggest a deep symmetry between action and perception. The construction of artificial intelligences founded on these principles has been largely absent. In this project we will construct control systems for artificial agents based on the FEP. In the best tradition of artificial intelligence research this work will not only shed light on the underlying assumptions of the FEP but will also inspire innovative new approaches to robotics. In particular the work promises significant fusion with methods of probabilistic robotics, a current industry standard. The student will gain experience in current hot topics in robotics including information theory, control theory, agent based approaches and exposure to conceptual advances in the cognitive sciences.

Detailed Description

The central deliverable of this project will be the practical and conceptual development of control systems based on the FEP for simple wheeled agents engaged in minimally cognitive task in software (these will include simple self-localisation and active categorical perception tasks) .The formalism of the FEP inherits directly from methods of variational Bayes (itself an extension of expectation maximization) developed in machine learning as a way to practically implement Bayesian inference [1]. The technical aspects of the FEP are widely documented [2] and both EASy group and Sackler centre are developing a growing expertise on these issues [3,4]. Consequently the first stage of this project is eminently achievable.

The central theoretical component will be to examine the feasibility of the  assumptions of the FEP but also to address a tension at the heart of AI concerning the role of models in cognition. To what extent does utilising internal models demand the capacity for representation? Is this framework naturally at odds with strong enactivism?

For engineering this research has the potential to broaden the scope of existing probabilistic robotics (PR) methodologies [5]. While PR architectures involve Bayesian inference and probabilistic control unlike FEP these two aspects are kept as relatively independent processes restricting the operational scope of agents to slow changing or static environments. By allowing coupling between action and perception within the same probabilistic framework the FEP has the potential to extend PR control to dynamic tasks in changing environments.

In addition a central challenge of modern robotics is the reality gap problem, i.e., how can we develop and test quickly and efficiently in simulation but transfer solutions to physical hardware. The FEP suggest a  possible solution to this because the internal models necessary for control systems are incomplete and developed for noisy environments potentially transferring between software and hardware with much greater ease. The student will be encouraged to explore these questions with standard robotics architectures, e.g., k-junior, lego mindstorms, to reduce lead time.

References

[1] Hinton, G and von Camp, D, Proceedings of COLT-93 (1995) [2] Friston, K., Trends in cognitive sciences  (2009)[ 3] Buckley, C. L.  et al., in prep (2014) [4] Seth, A. K., Cognitive neuroscience (2014) [5] Thrun, S. et al., Probabilistic robotics. MIT press, (2005)

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: