What Neuroscientists Don't Yet Know
Neuroscience is the study of how the brain works. If we knew how the brain worked we could replicate it. IBM's TrueNorth project had at one point on it's roadmap plans to build a system with 4 times the number of neurons as the human brain. The problem is not one of hardware, but one of software; we don't yet understand how the brain is wired and thus how it works. If we understood the algorithms and data structures used by the brain we could mimic them in silicon and software. Such mimicry need not be an exact replica, but need only capture the essential details as is the case with NeuralNetworks. Neuroscience is thus important as it helps determine how much hardware might be required to replicate human brain level performance, and it provides one path towards the construction of Artificial General Intelligence (AGI). With this in mind, here is a run down of what neursocientists don't yet know:
What is the neural code? Surprisingly, how information gets encoded in the brain is not presently known. There are a number of coding possibilities:
- rate coding - in which the frequency of neural firing corresponds to the strength of an informational signal
- temporal coding - in which the time between firings of different neurons carries information
- population coding - in which the firing of a large population of neurons is responsible for carrying information; if the brain uses a population code it may be possible to simulate it using an AGI architecture with far far less hardware than otherwise would be the case
- sparse coding - in which the firing of different small subsets of neurons encodes information
- How is time represented in the brain? How is sequentially ordered information, such as visual and auditory perception, language processing, and motor behavior, represented and produced by the brain?
How does the brain mediate dynamic communication between different brain functional regions? Brainwaves: large scale rhythmic activity of neurons in the brain at different frequency bands in the range 1-70 Hz are well known. See neural oscillation in Wikipedia. What is now emerging is the possibility that these different frequency bands may be used by the brain to carry signals that co-ordinate different brain functional regions. See for instance Visual areas exert feedforward and feedback influences through distinct frequency channels. This needs to be better understood.
- What are all the feedback connections for? The brain includes a very large number of feedback neural connections. It currently appears their purpose relates to attention and expectations. A deeper understanding of feedback connections could enhance neural network derived AGI architectures.
Just about anything related to how the brain develops. If intelligence is only possible after a prolonged developmental period this will greatly slow down exploring the space of possible AGI architectures. In development of the nervous system in humans Wikipedia differentiates between early stage brain activity independent mechanisms involving cell differentiation, migration, and axon guidance, and later stage activity dependent mechanisms that result in the formation of new synapses as well as synaptic plasticity underlying learning and memory. Since early stage development is believed to be occur independent of the effects of sensory inputs and brain activity it is possible to essentially skip this step and generate AGI architectures that need only be trained through activity dependent mechanisms. This is the approach taken by neural network based approaches to AGI, where a human designs the architecture and stochastic gradient descent or some other technique is used to set the synaptic weights, but the problem is the resulting architectures might not be capable of expressing human like intelligence. If more was known about how the brain develops and ultimately how it is wired at birth, more refined neural network based architectures might be possible.
- How is the brain so fast given that neurons are so slow? Neurons are slow, firing for anywhere from 1ms to 100ms. Despite this, the brain is relatively quick, being able to correctly respond to inputs in 200-400ms. This speed imposes significant constraints on the neural code and algorithms that must be used by the brain.
What is the effective human brain cortical neuron fan in/out? Divide the estimated number of neurons by the estimated number of synapses and you get an estimate of several thousand synapses per neuron. But there may be multiple connections to the same neuron, and in the absence of population coding a solitary synapse is unlikely to be able to drive a connected neuron. The effective fan in/out of cortical neurons in terms of unique neurons in the human brain is unknown. One stochastic estimate for the rat came up with 255 connections per neuron, see Reconstruction and Simulation of Neocortical Microcircuitry Why does this matter? A unit in a convolutional neural network may only receive 25 inputs. IBM's TrueNorth spiking neural chip has 256 synapses per neuron. If the effective fan in/out in the brain is larger this suggests AGI architectures might need to search over models with higher connectivity, along with an associated slowdown in model performance and an increase in search time.
Do dendrite branches function as active processing elements? Conventional neural network derived AGI architectures model neurons as point like integrators, An alternative architecture proposed by Hawkins of Numenta is that dendrite branches function as pattern recognizers. See Why Neurons Have Thousands of Synapses, a Theory of Sequence Memory in Neocortex.
How does episodic memory work? Neural networks require repeated exposure to inputs in order to remember. The brain can remember based on a single episode. The hippocampus and prefrontal cortex are believed to be involved in episodic memory. Better understanding how they interact and how the brain remembers from a single episode could define new AGI architectures.
- How does working memory work? There is a gap between working memory as implemented in recurrent neural network architectures with long-short-term-memory and how it is likely implemented in the brain through the integration of the prefrontal cortex, posterior parietal, and inferior temporal areas. Insight into how working memory works could set neural network derived AGI architectures off searching in new directions.
- What is the relationship between working memory and long term memory? How is information transferred from working memory to long term memory, and how are long term memories retrieved for use in working memory?
How does attention work? The brain prioritizes and focuses on a small subset of its sensory input. Exactly how attention works isn't know. Performing the same trick with neural network derived AGI architectures has the potential to reduce their cost or speed them up by an order of magnitude.
- What is the relationship between attention and working memory? Memory is limited, and so attention is important in determining what things to encode, and conversely past memories must play a role in determining what to attend to. This interplay isn't understood by neuroscientists.
- How does the brain remember new information without forgetting old information? Today neural networks when trained on a new task forget how to perform the old task. Understanding this could advance neural network derived AGI architectures. Sleep may play a role here, but this is far from settled.
- How does the brain solve the invariance problem. The visual system recognizes objects independent of their size and position. This is an interesting technical question, although one that might not need to be answered. Neural networks are capable of the same feat without us needing to understand how they do it.
How does the brain learn from a single instance? The brain can learn and generalize from a single instance. The question here may or may not be a restatement of how does episodic memory work and how does recognition occur independent of size and position. Understanding this could drastically speed up the time it takes neural networks to learn, and thus the speed at which different possible AGI architectures can be tried out. There are some one-shot machine learning algorithms already. See for instance Matching Networks for One Shot Learning.
The details of synaptic plasticity. Hebbian synaptic plasticity is one of the corner-stones of how the brain remembers. Much is known about how it works, but many details across the full range of time spans and how it is modulated remain unknown.
What is the function of a cortical column? The neocortex is made up of cortical columns, each containing perhaps 50,000 neurons. These cortical columns are often believed to reflect a functional role, but the computation performed by a cortical column remains unknown. If it turns out cortical columns are performing some relatively simple computation, this could significantly accelerate the emergence of AGI.
What does the thalamus do? The thalamus is traditionally seen as a relay between different parts of the brain. But is at a simple relay, or does it play an active role in deciding what signals to relay, and if so how is it controlled? It has even been suggested that the thalamus plays an important role in high level cognition Thalamic control of human attention driven by memory and learning.
How do the cerebellum and basal ganglia work? The cerebellum is believed to play a key role in performing motor control, namely predicting the sensory consequences of our motor actions. It has little to do with AGI. The architecture is relatively simple and it is known in considerable detail. Yet we don't know exactly how the cerebellum works. If we can't figure out the cerebellum do we have any hope of figuring out the far more complex neocortex, which is widely believed to be responsible for cognition? And do we need to understand how something works in order to be able to replicate it efficiently?
How is reinforcement learning incorporated into the the brain? Dopamine is widely seen as a temporal-difference learning signal in the brain. But how the basal ganglia respond to dopamine isn't well understood. Deep reinforcement learning is an emerging field, and it seems likely to mature without needing further insights from neuroscience.
- How does the brain perform hierarchical problem decomposition? Tasks involve sub-tasks, and the brain seems to effortless transition from one sub-task to the next effortlessly managing any contingencies. Teaching neurally inspired AGI to perform hierarchical problem decomposition has a way to go.
- What role if any do glia and astrocytes play in cognition? The traditional view is it isn't necessary to incorporate bidirectional astrocyte-neuron communication in AGI architectures.
- What is the role of spontaneous or resting state activity? How does it influence thoughts, behaviors, and task performance? This may or may not be relevant to neuron inspired AGI.
- What is the relationship between structure and function? How are anatomical and functional brain networks related? A single anatomical area may house multiple distinct functional networks.
- How do the two hemispheres of the brain integrate and work as a unified whole? This is probably not relevant to AGI.
- How is the brain modulated? Most neurotransmitters exert their influences over very short distances, but some appear to exert their effects more broadly. This doesn't appear well understood.
- What is the neuronal basis of psychiatry and psychology? This has many implications for AGI. Without any understanding of this, it would be just as likely for a neurally inspired AGI to exhibit depression or mania, as to be more balanced.
- What is consciousness and how does it emerge? Neuroscientists are a long way away from being able to answer this question. It may not need not be answered to develop AGI.
- What is the nature of free will? Is it real or an illusion? Answering this question is probably not necessary for developing AGI.
Clearly it will be a long time before we fully understand how the brain works. There are a lot of questions. Not all of them may need to be answered to achieve AGI. One thing is clear, the search space of possible neural network derived AGI architectures is very large. Most neurally inspired AGI projects can only attempt to pick answers to one or two of the above questions. This suggests that if something rivaling the complexity of the brain is necessary for AGI, it will be a long time before AGI is achieved. The question then is, how much of this complexity is really necessary for AGI?
How much of this complexity is necessary for AGI?
The argument against complexity being necessary for AGI is evolution works through a process of local optimization that typically leads to globally inefficient solutions. The entire reptilian brain might be capable of being replaced by a tiny piece of neocortex, but isn't because there is no evolutionary advantage to doing so, and a lot of ancillary processing baggage has come to rest on the reptilian brain, that would then need to be solved some other way. In short, evolution produces ugly-hack type solutions.
The argument for complexity being necessary for AGI is that if we seek to build a machine that displays human level abilities across the full range of human intellectual domains then AGI is likely going to need to possess the full range of human functionality, including functionality that might not be optimal from a pure intelligence perspective, but are necessary for the AI to fit in and excel in the present human defined world.
A middle ground might be that it will be relatively easy to develop simple AI architectures that greatly outperform humans in many domains but fail in some domains. But to create true AGI (which by definition is competitive with humans across all domains) would require much more complex and more difficult to develop AGI architectures that more closely mimic the human brain.