Adaptive cognition with context-aware and flexible neurons for the next generation of artificial intelligence
Neuromorphic computing mimics the structure and functional principles of the human brain in hardware and hopes to replicate the brain's intelligence.
In general, a significant and sought-after ability of the brain is its adaptive intelligence, which enables the brain to regulate its functions "on the fly" to effectively respond to changing circumstances, situations, goals, and rewards.
A concrete example of this adaptive intelligence is situational awareness, which is known to be a sign of cognitive or higher-level intelligence.
Contextual awareness is the ability of an agent to adjust its response to a situation or change its decision in a situation based on a basic situation marked "context".
It allows the brain to make complex, high-level decisions while considering multiple factors, understand the situation, respond to new stimuli, and make predictions from small data sets.
Another adaptive cognitive function found in the brain is cross-frequency coupling (CFC), which enables cognitive control of low-level neural signals through high-level knowledge.
One flavor of CFCs is phase-amplitude CFCs, in which global, low-frequency neural oscillations are phase-controlled or adapted to the amplitudes of local, high-frequency neural oscillations, where neural oscillations are rhythmic activities in the brain that occur at all levels of neural tissue.
CFC enables the brain to dynamically reorganize its internal network, integrate various functional systems, and control internal signaling.
The third adaptive capacity, called feature binding, enables the brain to combine different features of perceived objects into a coherent whole and integrate multimodal information to establish coherent representations of the external world.
It is thought that in the brain, the phases of global, low-frequency neural oscillations combine features by controlling or adapting the frequencies of local, high-frequency neural oscillations.
It is well known that feature binding is important for visual cognition and is related to consciousness.
The realization of these cognitive abilities in neuromorphic computers could lead to artificial intelligence (AI) with transformative implications.
The adaptive capacity of human intelligence described earlier derives from the adaptive capacity of biological neurons as well as the plasticity of biological synapses, mediated by the collective dynamics of neural circuits.
Adaptive neurons refer to the properties of neurons, including their state, transfer function, and state space, that can be changed or modulated "on the fly".
Biological neurons are adaptive through neuroregulation, that is, the regulatory input to neurons is used as a control signal, changing the characteristics of neurons and adapting their functions to different situations.
Through neuromodulation, the brain can change the amplitude and frequency of neural oscillations.
This adaptive capacity of biological neurons, especially neural regulation of neural oscillations, is essential for many additional cognitive processes, including information transfer, decision-making, memory, object representation, visual perception, and attention.
In addition to neuroscience, the importance of adaptive neurons is well recognized in software-based artificial neural networks (ANNs) and machine learning.
It is worth noting that the use of adaptive neurons in ANNs can accurately represent a continuous function using only one hidden layer.
This means that, in general, adaptive neurons are more expressive than non-adaptive neurons in representing continuous functions.
In software-based ANNs, neurons with adaptive states and adaptive transfer functions have been implemented and have been shown to achieve better classification performance than traditional non-adaptive neurons under smaller ANN architectures.
As a result, hardware-based adaptive neurons can enable cognitive abilities that were previously unattainable, with far-reaching applications.
The context-aware computing shown here can have a comprehensive impact on human-robot collaboration, personalized medicine, advanced manufacturing, and education.
In addition, feature binding enables multimodal information fusion, which can transform autonomous vehicles, neuroprosthetics, wearable health technologies, agriculture, and climate control.
More broadly, hardware-based adaptive neurons can alleviate several key challenges facing contemporary ANNs.
Modern ANNs require large amounts of training data, unsustainably large energies, and chip area when implemented on CMOS circuits, and are highly specific in their applications.
IN CONTRAST, ADAPTIVE NEURONS LIKE T-SKONE CAN ALLEVIATE THESE CHALLENGES, ENABLING ANNS TO LEARN FASTER, BE COMPACT, ENERGY EFFICIENT, FAULT-TOLERANT, AND ENABLE A WIDER RANGE OF ARTIFICIAL INTELLIGENCE.
Bibliography:
[1] Neftci E, et al. 2013. Synthesizing cognition in neuromorphic electronic systems. Proc Natl Acad Sci. 110(37):E3468–E3476.
[2] Grossberg S. 2021. Toward autonomous adaptive intelligence: building upon neural models of how brains make minds. IEEE Trans Syst Man Cybern Syst. 51(1):51–75.
[3] Dayan P. 2008. Simple substrates for complex cognition. Front Neurosci. 2(2):255–263.
[4] Dey AK. 2001. Understanding and using context. Pers Ubiquitous Comput. 5:4–7.
[5] Schilit B, Theimer M. 1994. Disseminating active map information to mobile hosts. IEEE Netw. 8:22–32.
