Nikkei Electronics Asia -- June 2009
Features
ASIMO Responds to Brain Activity with 90% Accuracy

Honda Research Institute Japan Co Ltd (HRI) of Japan, a subsidiary of Honda Motor Co Ltd of Japan, together with Advanced Telecommunications Research Institute International (ATR) of Japan and Shimadzu Corp of Japan, has developed brain-machine interface (BMI) technology capable of detecting, with a non-invasive brain activity sensor and with high precision, thoughts picturing actions.

The system achieved 90.6% accuracy in determining which one of four predetermined actions was thought of by measuring brain activity and analyzing the results. 4-option performance of about 60% has been announced in the past, but this is said to be the first time that a success rate of 90% or higher has been announced. The improvement was achieved through a combination of two brain activity measurement methods, and a proprietary discrimination method developed by ATR called sparse logistic regression (SLR).

Thought Control of ASIMO

At the press conference a video of Honda Motor's ASIMO humanoid robot being remotely controlled using BMI technology was shown. The operator wore a special helmet-type device and thought of one of four possible motions - right hand, left hand, feet and tongue. These thoughts were analyzed using BMI technology, and ASIMO performed the requested operation accordingly (Fig 1). For example, when the operator thought of moving the right hand, ASIMO interpreted the thought accordingly and raised its right hand.

In general, BMI technology determines the operator's intention based on information from inside the brain, which means that it is theoretically possible to create a user interface (UI) that does not require movement of hands, feet or eyes, for example, or speech. There is hope that this UI could be used by people with physical disabilities (hands or feet, for example), but HRI, ATR and other groups also believe that it could be invaluable even for people without such disabilities. For example, it might be possible to open your car door just by thinking about it when your hands are full of packages, or communicate with other people by thinking in situations where spoken conversation is difficult, such as in crowded trains.

The current implementation of BMI technology, however, only determines which of four commands has been issued, and the corresponding ASIMO motions are pre-programmed. It is not possible for the operator to think of an arbitrary motion, and have it reproduced by ASIMO. In addition, practical application as a UI would require a shorter delay from processing like measurement and recognition, and improved realtime performance. The delay is currently about seven seconds, making utilization as a UI difficult. A source at HRI comments, "We are still in the basic research stage, and not considering practical application any time soon."

Spatial-Temporal Precision

Accuracy was improved thanks to two key innovations. The first was combining two different methods of measuring brain activity: near-infrared spectroscopy (NIRS), which measures changes in cerebral blood flow by sensing near-infrared light reflected from the scalp, and electroencephalography (EEG), which measures changes in electrical potential on the scalp.

NIRS is unaffected by electromagnetic noise and therefore achieves a high spatial resolution, with measurement possible in units of several mm to several tens of mm. Unfortunately, it has a relatively long delay time, and low temporal resolution. EEG, on the other hand, has excellent temporal resolution (several ms), but poor spatial resolution because it measures scalp potential and not direct brain internal activity. The two methods were combined to resolve the weak points of both (Fig 2).

The other innovation used to boost accuracy was the adoption of a new method of analyzing measurements. NIRS, EEG and other measured data exhibit a different waveform for each measurement, even if the operator thinks of the same motion, such as "right hand". As a result, the feature values related to "right hand" must be extracted from a massive volume of measurement data. ATR announced a new method of extracting feature values and making discriminations in 2008: SLR, which is used to analyze functional magnetic resonance imaging (fMRI) data. This method was utilized in the newly developed BMI technology.

SLR extends conventional logistic regression* to include a Bayesian inferential framework, implementing dimensional compression of the feature value vector simultaneously with weighted modeling (inference) for discrimination. The technique is highly effective in situations where there is a large volume of extraneous feature value data, with a large number of feature value vector dimensions, as is the case with brain measurements. The technique sets the weighting parameter to zero for unneeded feature values during linear discrimination, extracting only a small (sparse) volume of feature value data related to the actual judgment. This makes it much easier to boost the precision of the BMI.

Support vector machine (SVM) is widely used as an analysis technique, but according to Mitsuo Kawato, director, ATR Fellow of the Computational Neuroscience Laboratories at ATR, "SVM is a pretty old technology for us. It just doesn't offer sufficient sparseness." SVM is designed to extract specific samples, and generally it does not act as feature extraction, and so it offers no functions capable of compressing dimensions as found in SLR.

by Tomonori Shindo

References:
1) Yamashita, O, et al, "Sparse estimation automatically selects voxels relevant for the decoding of fMRI activity patterns," NeuroImage, vol 42, 2008, pp 1414-1429

* Logistic regression: A type of linear discrimination analysis producing only a binary response (such as yes/no).

MOST POPULAR ARTICLES

24 Hours | 7 Days