3.1 Ecology

Whereas the cognitive sciences have long attended to information processing while ignoring energy, the field of ecology has long attended to metabolism while neglecting information. O’Connor et al. (Reference O’Connor, Pennell, Florian, Blake and Carlos2019, 3) reorient ecology by proposing five principles that integrate information into ecological understanding.

Principle 1: Information is a fundamental feature of living systems, and therefore, also of all ecological systems.

Principle 2: Syntactic and semiotic information interact in feedbacks, with energetic processes and material cycles, to influence structure, function and organization in ecological systems.

Principle 3: Information processing requires energy and materials; therefore, supply of energy and materials and thermodynamic constraints can limit information processing

Principle 4: Information processing allows components of living systems to measure the environment and their own state and to measure the relationship between their state and past and expected environments.

Principle 5: Information processing systems are linked within and across scales of biological organization.

Principles 2 and 3 are the most relevant to energy and substrate independence. O’Connor et al. use the terms “syntactic” and “semiotic” as follows:

Both kinds of information relate to energy. Syntactic information in the form of non-random organization is made possible by work done with energy in the environment. Spatial and temporal arrangements such as bird songs require energy expenditures to produce them. For semantic information, the acquisition of meaning through perceptual-motor interactions with the world requires energy in the form of light, sound, or muscular motion.

O’Connor et al. (Reference O’Connor, Pennell, Florian, Blake and Carlos2019) stress the interdependence of information with energy and material systems. Energy is required to create, maintain, and process information. Thus, the new emphasis on information in ecological theory ties it closely with considerations about energy, providing support for my premise that information processing is dependent on energy. The kinds of energy used in biological systems such as sense organs and muscles is dependent on their material constituents. For example, hearing depends on material structures of the outer, middle, and inner ear that gain their energy from glucose fueling for cells. Movement depends on similar fueling for muscle cells.

Metabolism is the sum of chemical reactions that sustain life, including conversion of food to energy, the conversion of food and fuel to structures such as proteins, and the elimination of wastes. The principles of O’Connor et al. (Reference O’Connor, Pennell, Florian, Blake and Carlos2019) indicate the relevance of information to metabolism because organisms use representations with syntactic and semantic properties to help them in finding food that fuels metabolism. Energy and information are interdependent in that an organism needs energy to acquire and use information, but it also uses information to acquire energy by gathering food. All organisms and hence the ecologies in which they operate are therefore subject to balances between energy and information. Information is not just structured data, code, or text, but must also be stored, manipulated, sent, and received in processes that require energy.

3.2 Brains

Brains are often considered to be information processors, but they are also notable as energy users. A 3-pound brain in a 150-pound body has only 2% of its weight but uses around 20% of its energy (Raichle and Gusnard Reference Raichle2002). The brain has around 86 billion neurons whose operation requires continuous supplies of glucose. Billions of neurons build up electrical charges and send electrochemical signals to approximately 10,000 neurons often more than 100 times per second. The expended energy in each neuron needs to be replaced by blood transport to neurons and glia cells that support them. When blood supplies are blocked by mechanism failures such as strokes, heart attacks, and asphyxiation, neurons die.

The correlation between brain size and intelligence is imperfect, as elephants and some species of whales have brains with more volume and neurons than humans. But more neurons usually means more intelligence, especially when the neurons operate in brain areas associated with complex problem solving such as the prefrontal cortex. Adding neurons, however, poses an evolutionary problem because of their intense energy needs. For brains to get bigger, food acquisition needs to increase despite the limitations of environmental ability, number of hours available for foraging, and energy cost of acquiring food.

Susanna Herculano-Houzel (Reference Herculano-Houzel2016) describes the steady increase in sizes of primate brains from lemurs to gorillas but notes the dramatic increase in human brain size that occurred in Home erectus around 1.5 million years ago. Following Wrangham (Reference Wrangham2009), she attributes increased brain size to the invention of cooking, which makes meat and plants much easier to digest and energy intensive. Because cooking dramatically increased energy efficiency, humans could evolve larger brains without having to spend more time and energy searching for additional food supplies.

Quantitative increases in human brain size led to qualitative tipping points in human intelligence with emergent capabilities not found in other animals. Around 50–100,000 years ago, large human brains became capable of recursive thought that provided capabilities for language, art, and tools that make tools (Corballis Reference Corballis2011; Coolidge and Wynn Reference Coolidge and Wynn2018). Recursion requires the interaction of large numbers of neurons capable of representations of representations (Thagard Reference Thagard2021). Without the efficiency of cooking and the development of large brains, humans would not be capable of the energy/information balance that eventually made human civilization possible. Less than 15,000 years ago, the invention of agriculture provided sufficient food sources that led to dramatic increases in the number of humans as well as the development of productive practices such as writing, counting, and using complex machines.

Thus, the evolution of human intelligence has been energy-dependent in several ways. Acquisition of sufficient food was essential to evolution of larger primate brains and got a large boost from the invention of cooking. Changing energy availability through cultivating crops and domesticating animals enabled the formation of larger human groups such as cities and divisions of labor necessary for many cultural advances. Moreover, the ways in which humans use energy in information processing is heavily dependent on the particular kinds of matter that operate in human brains – the intense but efficient operation of billions of glucose-supplied neurons.

For organisms in general and for human brains in particular, these considerations show that information processing depends on energy, which depends on the material substrate. Here “depends” means that information acquisition, processing, and application are causally affected by mechanisms with particular kinds of parts, interactions, and energy sources. Without such mechanisms, information processing would not be effective and efficient.

4.1 Neuromorphic computing

Carver Mead (Reference Mead1990) proposed the development of neuromorphic (brain-like) computers to deal with two potential limitations in computing technology: speed and power consumption. The speed of computers has increased exponentially since the 1960s in accord with Moore’s law that the number of transistors on a chip doubles every two years. But physical limits such as problems of heat dispersion put limits on how much smaller transistors can be built, so computer engineers must look to parallelism as an alternative way of increasing computers’ speed.

Most computers today are built with multiple “cores,” which allow parallel processing, but the number of cores is usually small, e.g., 4 or 28. In contrast, the brain has 86 billion neurons operating in parallel, which allows it to outperform current computers whose chips do trillions of operations per second in contrast to the approximately 100 or so spikes per second common in neurons. Companies like Intel and IBM have developed neuromorphic chips with more than 100,000 neuron-like processors operating in parallel. IBM’s TrueNorth chip has one million neurons and is capable of 46 billion synaptic operations per second while using only 70 milliwatts (Akopyan et al. Reference Akopyan, Sawada, Cassidy, Alvarez-Icaza, Arthur, Merolla, Imam, Nakamura, Datta and Nam2015). Intel’s Loihi chip has 128 neuromorphic cores each with 1,024 neural units (Davies, Reference Davies, Narayan, Tsung-Han, Gautham, Yongqiang, Sri Harsha, Georgios, Prasad, Nabil and Shweta2018). Schuman et al. (Reference Schuman, Potok and Patton2017) survey neuromorphic computing.

Besides speed, the major motivation for neuromorphic computers is power consumption. Whereas a traditional supercomputer can require thousands of watts of power, the brain uses only around 20 watts. The efficiency of computing can be measured by considering the ratio of operations per energy, for example, inferences per joule, neural spikes per joule, or (in supercomputers) floating point operations per second per watt. More abstractly, information efficiency can be measured in bits per joules.

Early experimental results find that neuromorphic computers are indeed more energy efficient than digital computers (Martí, Rigotti, Seok, and Fusi Reference Martí, Rigotti, Seok and Fusi2016; Blouw, Choo, Hunsberger, and Eliasmith Reference Blouw, Choo, Hunsberger and Eliasmith2019). There are several reasons for such advantages. Neuromorphic machines do much of their computations locally with artificial neurons connected to ones that are physically close to them, requiring much less long-distance transmission of information over wires. Algorithms for learning and memory operate locally so that data do not need to be widely broadcast. Processing elements in neuromorphic computers only emit spikes when required, whereas CPUs are constantly active. Parallelism reduces the number of steps to transform an input into an output. Finally, distributing computation across many processors reduces heat generation so that less energy is required for cooling. Hence, neuromorphic computing is attractive for applications such as biomedical devices where low power consumption and real time operation are desirable.

There are also time-related reasons why neural computers can be more efficient than traditional ones. Neural spiking may not operate as rapidly as traditional silicon chips, but it can operate in synchrony with inputs from processes in the world without any external clock. Time represents itself because of the correspondence between changes in the world and changes in the brain. Neural operations occur in step with changes in the world on the same time scale, although most current neuromorphic chips do not exploit this feature.

Neuromorphic computing illustrates the same energy-information tradeoffs that occur in biology. In general, the more syntactic and semantic information processed by a computer, the more energy it uses. But neuromorphic computers require less energy to perform the same task, so their information/energy ratio is better. As with biology, information processing depends on energy, and energy depends on material substrates.

4.2 Machine learning

Artificial intelligence has made major progress since 2012 because of powerful applications of a neural network technique called deep learning, which has led to spectacularly effective applications such as face and speech recognition (LeCun, Bengio, and Hinton Reference LeCun, Bengio and Hinton2015; Goodfellow, Bengio, and Courville Reference Goodfellow, Bengio and Courville2016). The AI company DeepMind has combined deep learning with reinforcement learning to produce amazing successes such as the world’s best Go player (Silver et al. Reference Silver, Thomas, Julian, Ioannis, Matthew, Arthur, Marc, Laurent, Dharshan and Thore2018). DeepMind researchers can afford to have their Go player improve by playing itself millions of times, but the resulting information is gained at the expense of a large amount of electrical energy. Schwartz, Dodge, Smith, and Etzioni (Reference Schwartz, Dodge and Smith2019) describe how the computations required for deep learning have been doubling every few months. Strubell, Ganesh, and McCallum (Reference Strubell, Ganesh and McCallum2019) report the financial and environmental costs of training neural networks for natural language processing.

The rapidly increasing use of electricity for deep learning and other intensive kinds of computing such as cryptocurrency is concerning for reasons that go beyond corporate expense. Standard ways of generating electricity, such as gas power plants, produce large amounts of greenhouse gases that contribute to global warming which threatens to change the Earth’s climate in disastrous and irreversible ways in a matter of decades. Reducing energy use for deep learning and other AI applications should therefore be a general social concern, and research is investigating generally how to make computation more energy efficient (Demaine, Lynch, Mirano, and Tyagi Reference Demaine, Jayson, Mirano and Nirvan2016).

In contrast to current machine learning, human brains learn from much smaller numbers of examples with energy-efficient processes that make local changes in the strength of synaptic connections. Brains have ways of making learning more efficient by considering relevant subsets of information in data such as ones representing causal relations that circumscribe what might be relevant to human goals. Deep learning has no understanding of causality, so it needs huge numbers of training examples to extract signal from noise. Thagard (Reference Thagard2021) examines what it would take to give computers a human-level capacity for causal reasoning.

Attempts are underway to make deep learning more efficient by performing it on neuromorphic computers (Esser et al. Reference Esser, Merolla, Arthur, Cassidy, Rathinakumar, Alexander, Berg, McKinstry, Timothy, Barch, Carmelo, Pallab, Arnon, Brian, Flickner and Modha2016). An alternative research strategy is to develop new algorithms that allow AI systems to learn from small numbers of examples using human methods such as abductive inference to causes, analogy, and taking into account background information about variability (Holland, Holyoak, Nisbett, and Thagard Reference Holland, Holyoak, Nisbett and Thagard1986). Applications of machine learning have been able to rely on abundant sources of electricity, but constraints such as global warming and portability suggest that AI researchers need to pay attention to energy/information tradeoffs. So deep learning developments point to further connections among inference, energy, and material substrates. Another area of current research concerns the prospects of quantum computing for increasing energy efficiency (Ikonen, Salmilehto, and Möttönen Reference Ikonen, Salmilehto and Möttönen2017).

So deep learning and neuromorphic computing both show the relevance of energy to understanding and evaluating information processing in computers. As for biological systems, efficient use of energy requires attention to material substrates.

5.1 Church-Turing thesis

In the 1930s, Alonzo Church and Alan Turing independently produced mathematical characterizations of effective computation. Church’s lambda calculus and Turing’s abstract machine turned out to be equivalent to each other and to Kurt Gödel’s characterization of recursive functions. This convergence is taken to support the plausibility of the Church-Turing thesis according to which every effective computation can be carried out by a Turing machine (Copeland Reference Copeland2017). Tegmark (Reference Tegmark2017, 65) says, “This fact that exactly the same computation can be performed on any universal computer means that computation is substrate independent.”

The Church-Turing thesis might be taken to provide support for substrate independence because it shows that computation does not depend on a particular kind of computing method. But all three mathematical formulations (Church, Turing, Gödel) ignore time, space, and energy, so the Church-Turing thesis says little about real-world computational devices. Two computers might be equivalent in what they compute but differ markedly in how long they take to do so and how much energy they consume, as we saw in the difference between traditional digital computers and neuromorphic computers. In real life, bits are not free. The mathematical plausibility of the Church-Turing thesis provides no support for the doctrine of multiple realizability, because it does not address the question of how computation can be realized in real physical systems bounded by time and space as well as energy.

5.2 Gradual replacement

Another argument for substrate independence was used by David Chalmers to support the feasibility of uploading conscious minds into computers:

If the parts of one system can be gradually replaced with very different parts without loss of function, then the actual composition or substrate does not matter. Like philosophical thought experiments in general, the brain replacement exercise underspecifies the scientific facts in ways contrived to reach a desired conclusion (Thagard Reference Thagard2014, Reference Thagard2019b). Parts such as neurons do not operate in isolation, but rather in mechanisms where connected parts interact to produce regular changes. Whether the different kinds of parts perform the same function depends not just on the parts but also on the connections and interactions, both of which require energy. If the energy uses of the replacing combination of parts are very different from the energy requirements of the original combination, then it is unlikely that the gradual replacement will work.

Consider the mediocre history of artificial heart replacement. The Jarvik mechanical heart was first implanted in a human in 1982, and many variants have since been produced. Decades later, however, artificial hearts are only used as temporary replacements because of problems that include immune system rejection, durability, and power supply (McKellar Reference McKellar2018).

Similarly, gradual replacement of groups of neurons with silicon chips has substantial obstacles of space, time, and energy. Silicon chips are much larger than neural groups, so they would have difficulty connecting with them within the confines of the skull. Timing is very different in neurons and chips, even neuromorphic ones, so there are major difficulties in coordinating the interactions of chips and neurons. Finally, the problems of systematically replacing the biochemical energy sources of neurons with the electrical energy sources of silicon chips are large. In accord with the five characteristics I identified, use of energy has to satisfy constraints concerning limited supplies, costs of operation, consumption rates, efficiency, and tipping points. All of these are affected by the materials that use energy in different ways.

The gradual replacement story is appealing to those who already believe that substrates are not important to thinking and computation. But attention to the mechanisms that operate in different systems show that actual replacement is problematic because of constraints of space, time, and energy. So the abstract possibility of replacement does not support the plausibility of substrate independence.

The replacement argument encourages the hypothesis of multiple realizability that is often used to show that mental properties cannot be reduced to brain properties. Critics of the idea of multiple realizability have pointed out that cognitive neuroscience has in fact made substantial progress by connecting aspects of human cognition such as vision with neural mechanisms (Bechtel and Mundale Reference Bechtel and Jennifer1999; McCauley and Bechtel Reference McCauley and Bechtel2001; Thagard Reference Thagard1986, Reference Thagard2019a). In computer science, it has become increasingly evident that hardware and software are not independent, and that practical success requires that algorithms match hardware. The Church-Turing thesis and replacement argument fail to support substrate independence because of mounting evidence that information and energy depend on material details.

6.1 Endothermy

Polger and Shapiro (Reference Polger and Shapiro2016, 142) mention that endothermic (warm-blooded) animals must acquire large amounts of energy and therefore consume much more food than ectothermic (cold-blooded) organisms. There are indeed important relationships among metabolism, body-brain size ratios, and neuronal densities (Yu, Karbowski, Sachdev, and Feng Reference Yu, Karbowski and Sachdev2014). Larger brains with more neurons consume more energy, which is a challenge for organisms that need energy to keep themselves warm. Accordingly, ectothermic animals in tropical climates have larger brains than ectothermic animals in colder climates. A key factor in the development of large-brained endotherms is the rapid increase in the number of glial cells that are less energy-demanding than neurons.

These findings show that comparisons of psychological functions such as vision across different organisms cannot ignore energy considerations. Such functions operate in environments that include factors of external and internal temperature that interact with the neural mechanisms that carry out the psychological functions. Hence, determination of whether psychological functions are multiply realized needs to include energy assessments. Mammal and bird brains operate in warm-blooded bodies unlike the cold-blooded bodies of molluscs, reptiles, and fish, which can affect basic neural operations such as the balance between number of synaptic connections per neuron and strength of weights between neurons (Yuan et al. Reference Yuan, Huo, Zhao, Liu and Liu2018).

6.2 Information processing

Polger and Shapiro (Reference Polger and Shapiro2016, ch. 8) provide an extensive discussion of information-processing models in cognitive science. They recognize that such models have been a major motivation for functionalist accounts in the philosophy of mind: if human thinking is information processing, which can also be done by computers, then mental functions are multiply realized.

But what is information? Recent surveys in the Stanford Encyclopedia of Philosophy provide no general answer (Adriaans Reference Adriaans2018; Floridi Reference Floridi2015; Godfrey-Smith and Sterelny Reference Godfrey-Smith and Kim2016). Dictionary definitions of information as facts or knowledge are not informative. Floridi Reference Floridi2010 defines information as data that are well-formed and meaningful, but this is of limited use when dictionaries commonly define a datum as a piece of information and Floridi’s own account of a datum as something distinct is vague. The elegant probabilistic characterization of information by Shannon and Weaver (Reference Shannon and Warren1949) has little application to biological systems.

A better way of characterizing information is to specify the classes of mechanisms by which it operates, which include representation, collection, storage, retrieval, evaluation, transformation, sending, and receiving. All of these operate on different carriers of information that include utterances, sentences, gestures, pictures, patterns of neural firing, computational processes, and genetic structures made of DNA. I will illustrate the eight mechanisms using neural patterns.

Representation allows carriers to stand for situations in the world. For example, different groups of neurons are tuned, either innately or by learning, to different aspects of the world such as colors. Such tuning enables collection of new information by interaction with the world, for example, when the retina sends signals from the eye to the brain where storage takes place by generation of synaptic connections. For ongoing use, the information must be retrieved from storage, which the brain accomplishes by firing the neurons based on their synaptic connections. Whether a carrier of information is collected, stored, and retrieved depends on its relevance to an organism’s goals such as survival and reproduction, where relevance is assessed by an evaluation mechanism. In brains, evaluation is largely performed by emotional brain areas such as the amygdala, nucleus accumbens, and ventromedial prefrontal cortex. Information processing also includes making inferences that transform carriers into new and useful ones using mechanisms that range from learning by association to analogical reasoning. For many species of organisms such as humans, information is communicated between individuals, which requires mechanisms for both sending and receiving different kinds of carriers. Humans have neural processes for producing utterances as well as for interpreting the utterances of others.

All these mechanisms require energy to keep the parts working, to keep the parts connected, and especially to accomplish the interactions of parts that produce regular changes. Neurons need energy to survive, maintain synaptic connections with other neurons, and interact by processes of excitation and inhibition. Hence, determination of whether information processing is multiply realized must consider the extent to which energy constitutes a relevant similarity or difference. Because neurons and computer chips have mechanisms that use energy in very different ways, we cannot assume that abstract information processing supports functionalism.

Information processing requires attention to speed as well as inputs and outputs. If the eight information mechanisms operate too slowly, then an organism will not be able to accomplish its mental functions in its environmental context. For example, if an animal is too slow in its transformations of neural patterns, then it is not able to recognize food and predators fast enough to survive and reproduce, with resulting extinction. Fast processing is heavily dependent on energy in the same way that running consumes energy more rapidly than walking. Hence, debates about the multiple realization of information processing should take into account time/energy tradeoffs.

6.3 Identity theory

Polger and Shapiro’s critique of multiple realization arguments is in defense of their version of a mind-body identity theory, but Endicott (Reference Endicott2017) remarks that they do not advance specific mind-body identity claims. In contrast, Thagard (Reference Thagard2019a, b) specifies neural mechanisms for a host of important mental processes that include perception, imagery, concepts, rules, analogies, emotions, consciousness, actions, intentions, language, creativity, and the self. The requisite mechanisms are based on Eliasmith’s (Reference Eliasmith2013) Semantic Pointer Architecture, which pays close attention to the brain’s energy usage. Hence arguments for mind-brain connections can in part be energy-based.

At first glance, it might seem that the Semantic Pointer Architecture supports multiple realization because its large-scale simulations with millions of artificial neurons have largely been run on conventional computers (Eliasmith et al. Reference Eliasmith, Stewart, Xuan, Trevor and Travis2012). But Eliasmith and Stewart (Reference Eliasmith and Terry2020) explain why the most recent versions are being run on neuromorphic chips that mimic the energy efficiency of the brain. They estimate that an artificial neural network the size of a human brain would take half a gigawatt of power using today’s computers, equivalent to the power produced by a nuclear power plant. Use of Intel’s Loihi neuromorphic chip is far more energy efficient and makes increasingly large simulations feasible.

Polger and Shapiro (Reference Polger and Shapiro2016, 158) remark that connectionist neural networks can be built out of most anything, but considerations of energy and time show that multiple realization fails for them as well as for many other proposed systems. A neural network built out of toilet paper would be too slow, fragile, and energy demanding to be of any use. Future defenses of the richness of mind-brain connections can take into account ways in which the effectiveness of psychological functions depends on the energy-efficiency of the brain and neuromorphic computer chips. The use of specialized machines for brain simulations does not support multiple realization because there are still important energy differences between brains and neuromorphic computers, for example, in the role of glial cells and different kinds of neurotransmitters.

I have shown how energy considerations concerned with endothermy, information processing, and mind-brain identities support the rejection of multiple realization arguments for functionalism. The philosophical debate is ongoing, as serious responses to Polger and Shapiro (216) have been given by Aizawa (2017), Chirimuuta (Reference Chirimuuta2018), and others. Garson (Reference Garson2003) and Piccinini (Reference Piccinini2015) maintain that information and computation are “medium independent”, another term for substrate independent. I hope that future discussions will include close attention to energy requirements.