Throughout the history of mankind, we have continuously developed new tools such as letters, money, printing technology, and computers. Every time a new tool developed, our brains adaptively learn how to use it through experience and actively make use of it in our daily lives. This process arises from “plasticity” of the brain. One of the important aspects of plasticity is that the structure and function of the brain circuitry change in response to information obtained through the senses, such as the eyes and ears. Thanks to the plasticity, we are able to learn and remember things. In other words, human beings have not only developed science and technology, but have changed the way how to use their own brains according to the development of science and technology. This fact implies that the brain still has high potentials, but it cannot fully demonstrate that potential while being placed in our current “raw” environments. Therefore, by making the best use of AI, we would like free the brain from the constraints of the body and attempt to bring out its hidden potentials.

We will explore “symbiotic emergence” from a hybrid of the brain and AI, and develop a new dimension of perception, sensitivity, and cognition. The “symbiotic emergence” refers to the extension of intelligence that occurs as a result of interactions between the brain and various objects such as the environment, body, and machines, and our goal is to develop AI that can support such interactions. Specifically, we will work on the development of technologies to expand the interfaces and brain capabilities to support interactions between the brain and outside world, as well as technologies for elucidating the principles and mechanisms for understanding interactions and to collect and analyze information that contributes to this understanding.

The first approach to achieve “symbiotic emergence” between the brain and AI is to use an AI chip with built-in sensory sensors. Signals from the environment and body are detected by sensory devices. Based on the detected information, the AI chip sends stimuli to the brain to make it recognize information that it would not normally be aware of. For example, most Japanese people are not good at differentiating between the R and L pronunciations. Because we grow up hearing Japanese language, our ability to distinguish between R and L pronunciations was lost as a function. The pronunciations of R and L are physically different, but different signals must be sent from the ear to the brain; nonetheless the brain does not discriminate between them. Thus, we can design a AI chip that produce different electric signals depending on the pronunciation, so that the brain can recognize the difference. Repeated use of this device will induce “plasticity” in the brain and expand the ability to distinguish pronunciation.

Another approach is to have AI decode the brain information of individuals and then inform them about the results. In some cases, individuals are not able to utilize external information detected by the brain. Using AI to decipher such hidden assets and inform individuals can help them learn skills quickly, even those that would normally require prolonged periods of training to master. For example, we will use AI to learn skills that require a high level of judgment by integrating information obtained from past experiences and the environment, such as insight to determine a good or bad move by looking at a chess board, or an aesthetic eye when appreciating a painting.

This research attempts to transcend academic disciplines by engaging psychiatry, psychology, and pedagogy with neuroscience and information science. The questions to be addressed are: 1) What is intelligence?; 2) Can intelligence be extended?; and 3) How should the intelligence be utilized? In this research, we will approach these fundamental questions through the development and implementation of brain-AI hybrid technology. Our future goal is to augment humanity and develop next-generation intelligence with new values by implementing into the brain perceptions and superior cognitive abilities that animals and humans cannot innately perceive (e.g., absolute pitch, speed reading), as well as sensitivities and skills that cannot be acquired without long-term training (e.g., the aesthetic sense, insight, and inspiration).

Neurofeedback, which visualizes and self-regulates one's own brain activity, is effective in improving visual and auditory perception. However, there are some neural activity patterns that cannot be acquired by self-regulating neurofeedback, which is a limitation in expanding brain functions. To overcome this limitation, we hypothesized that brain functions can be maximally expanded by extracting latent information not used by the brain through machine learning and feeding it back directly to the brain as electrical stimuli. To test this hypothesis in an animal model, we set up two tasks that rats cannot naturally discriminate. One is heartbeat control and the other is speech discrimination. Both tasks are described below.
Physiological parameters primarily controlled by the autonomic nervous system, such as heart rate (HR), blood pressure, and body temperature, can be intentionally regulated through specialized training that provides real-time feedback to the individual. This biofeedback technology has potential clinical applications, but is also relevant to the practice of yoga and meditation, and to sports such as free diving and shooting. Despite its wide range of applications, however, the neural basis of brain-to-tissue control during biofeedback remains poorly understood. Inspired by our previous work, we therefore developed an experimental model of HR feedback using free-ranging rats to elucidate the neural mechanisms governing HR feedback training. Specifically, we stimulated the neocortex and medial forebrain bundle as HR feedback and reward, respectively. Rats learned to reduce HR within 30 min and achieved a reduction of approximately 50% after 5 days of 3-hr feedback training; the HR reduction persisted for at least 10 days after the 5-day training period, at which time the rats exhibited anxiolytic behavior and also had an increased number of erythrocytes in their blood.
　We then examined the neurophysiological activity and neural circuitry underlying the biofeedback-induced bradycardia. Bradycardia was prevented by inactivating anterior cingulate cortical (ACC) neurons projecting to the ventromedial hypothalamic (VMT) nucleus. ACC neurons projecting to the VMT exhibited theta-band field oscillations during operant training, and optogenetic stimulation of the ACC-to-VMT pathway with theta rhythm resulted in a HR decrease. VMT neurons receiving synaptic input from the ACC project to the dorsomedial hypothalamus (DMH), whereas DMH neurons project to the nucleus ambiguus (Amb), an autonomic center innervating the postnodal parasympathetic innervation of the heart. Our results show that ACC→VMT→DMH→Amb→heart projections are responsible for the top-down pathway of spontaneous HR regulation, and are the first in the world to reveal how biofeedback works. This achievement brings us one step closer to the old and new great mystery of how body and mind are connected.
The second is an experiment in which rats were asked to discriminate between English and Spanish. We used a speech synthesis model with 50 English and 50 Spanish phrases, each with the same voice quality. The rats were placed in a box with two nosepoke holes on either side and randomly presented with either English or Spanish. 500 trials per day for a week resulted in a language discrimination rate of 50%, which is no better than chance. In other words, the rats could not discriminate between the human languages English and Spanish. Therefore, we recorded local field potentials (LFPs) from the auditory cortex during language presentation and used a convolutional neural network (CNN) to identify which language was being presented from the LFPs; the classification accuracy exceeded 60%, which was significantly higher than the chance level of 50%.
We then fed the CNN responses back to the rats through electrical stimulation of the left or right somatosensory cortex (S1). Electrical stimulation of the medial forebrain bundle (MFB) was used as a reward. As a result, the rats were able to nose-poke the correct hole approximately 90% of the time. To determine whether learning was maintained when the feedback stimulation was turned off, we conducted a design with no feedback stimulation for 10% of the trials and found that the correct response rate was maintained even in the absence of stimulation. Our study showed that intelligence enhancement is possible through co-learning between the brain and artificial intelligence.　
If the technology of this study is extended to humans, it may hold promise as one of the new treatments for improving cognitive dysfunction. Furthermore, if this technology is applied to healthy people, it will enable them to acquire new concepts and values of things that they have not noticed before, and it is expected to contribute to improving the well-being of humanity.

Selected Publications
1. Norimoto, H., Makino, K., Gao, M., Shikano, Y., Okamoto, K., Ishikawa, T., Sasaki, T., Hioki, H., Fujisawa, S., and Ikegaya, Y. Hippocampal ripples down-regulate synapses. Science, 359:1524-1527, 2018.
2. Norimoto, H., Ikegaya, Y. Visual cortical prosthesis with a geomagnetic compass restores spatial navigation in blind rats. Curr. Biol., 21:1091-1095, 2015.
3. Ishikawa, D., Matsumoto, N., Sakaguchi, T., Matsuki, N., Ikegaya, Y. Operant conditioning of synaptic and spiking activity patterns in single hippocampal neurons. J. Neurosci., 34:5044-5053, 2014


※For more detail, please refer to the URL below.
https://beyondai.jp/contents/wp-content/uploads/2024/02/Research-report_Yuji-Ikegaya_202309.pdf